Update version numbers for TensorFlow 1.15.4

Update release notes for TensorFlow 1.15.4

Fix broken tests after cherrypicks

Remove import that is not needed

Fix import path

Cast away a const in intermediate API

Fix typo in macro

Fix import path

Solve leftover from merge conflict

Patch for TF 1.15.4

PiperOrigin-RevId: 332578058
Change-Id: I9727571d2f21476b10d8aa27c1b7176564b76ac9

Also add tests as they were lacking

PiperOrigin-RevId: 332566071
Change-Id: I44277578e26ff5fb3fdb0dcbba6e91b2ec3e7859

The `tensorflow::Shard` functions last argument must be a 2 argument function where both arguments are `int64` (`long long`, 64 bits). However, there are usages where code passes in a function where arguments are `int` or `int32` (32 bits). In these cases, it is possible that the integer truncation would later cause a segfault or other unexpected behavior.

PiperOrigin-RevId: 332560414
Change-Id: Ief649406babc8d4f60b3e7a9d573cbcc5ce5b767

The function argument in `Shard` must be a function of two `int64` arguments. However, we are passing in a function with two `int` arguments. Thus, for large workloads, these arguments get truncated from positive `int64` values to negative `int` ones, resulting in a buffer out of bounds write.

PiperOrigin-RevId: 332557334
Change-Id: I236c9a2e7f53580e520571da8ba941a3aa9fa0b5

The `printf` format specifier only allows `#`, `0`, `-`, `+` and space as flag characters. Others are interpreted as width/precision/length modifier or conversion specifiers. If a character does not fit into any of these sets `printf` just displays it.

Also add a test suite for `tf.strings.as_string`. Also fix the issue where the flag character was used only if width was specified.

PiperOrigin-RevId: 332553548
Change-Id: Ie57cf2a7c14d1a36097642794c14329db669bbba

In eager mode, session state is null.

PiperOrigin-RevId: 332548597
Change-Id: If094812c2e094044220b9ba28f7d7601be042f38

Without validation, we can cause a heap buffer overflow which results in data leakage and/or segfaults.

PiperOrigin-RevId: 332543478
Change-Id: Iee5bda24497a195d09d122355502480830b1b317

We already validated `NodeDef`s from a `GraphDef` but missed validating those from the `FunctionDefLibrary`. Thus, some maliciously crafted models could evade detection and cause denial of service due to a `CHECK`-fail.

PiperOrigin-RevId: 332536309
Change-Id: I052efe919ff1fe2f90815e286a1aa4c54c7b94ff

Also reorder fuzz generated test cases following f760f88b

PiperOrigin-RevId: 308339007
Change-Id: I11d825203964cf3397846c57fd4a6f458e8536f3

PiperOrigin-RevId: 308283636
Change-Id: Ib10849425de7d541d8dacfe4d0c709fbac9180b6

In Python, a list `l` of length `n` allows indexing with negative indices, `l[i]`. The only constraint is that `n + i` becomes positive. Code in `ResolveAxis` assumes the constraints and only checks it using a `DCHECK`. But the macro is a no-op in non-debug builds and that can result in reading from negative offsets (buffer underflows).

PiperOrigin-RevId: 332530683
Change-Id: I464e073fee618054ae3719a3679739007bb3f3bc

We check in `MatchingDim` that both arguments have the same dimensionality, however that is a `DCHECK` only enabled if building in debug mode. Hence, it could be possible to cause buffer overflows by passing in a tensor with larger dimensions as the second argument. To fix, we now make `MatchingDim` return the minimum of the two sizes.

A much better fix would be to return a status object but that requires refactoring a large part of the codebase for minor benefits.

PiperOrigin-RevId: 332526127
Change-Id: If627d0d2c80a685217b6e0d1e64b0872dbf1c5e4

A crafted TFLite model can force a node to have as input a tensor backed by a `nullptr` buffer. That is, by carefully changing the buffer index in the flatbuffer serialization, we can force the TFLite interpreter to consider a read-only tensor to be a read-write one and assume that there is an operator that has this tensor as output, writing to it and allocating memory before the tensor is used as input. If this does not happen, we get memory corruption.

PiperOrigin-RevId: 332524692
Change-Id: I57ef175152a29020af9ab041dc959e5631dce40f

If a model uses the same tensor for both an input and an output then this can result in data loss and memory corruption. This should not happen.

PiperOrigin-RevId: 332522916
Change-Id: If0905b142415a9dfceaf2d181872f2a8fb88f48a

With the previous change, there is no more need for two separate APIs. We would deprecate `GetOptionalInputTensor` in the future.

PiperOrigin-RevId: 332513386
Change-Id: Id7110271c25ebd6126ad8c82a493e37e0e0756b3

`GetInput`, `GetVariableInput` and `GetOutput` all fail to check for the case where `node->inputs->data[index]` is the special `kTfLiteOptionalTensor` value (-1) which then causes `context->tensors[node->inputs->data[index]]` to read from invalid memory location.

This fix makes `GetInput` and related return `nullptr` in those cases, asking the caller to check for `nullptr`. This is better than having `GetOptionalInputTensor` and `GetOptionalOutputTensor` (does not exist but could be added) as using the patched `GetInput` in error would be caught by a sanitizer test in the default optimized build (due to the `-fsanitize=null` option).

PiperOrigin-RevId: 332512190
Change-Id: Iabca54da2f2de02b6ece3c38b54f76d4277d689e

Buffers in the model are allocated globally, hence it makes sense to check for
their presence only once (O(1)) instead of on every subgraph (O(n)).

PiperOrigin-RevId: 323677724
Change-Id: I2da0c381093006828cc4c80f03dec8a917782861

Cherry-pick sqlite version bump