Fix negative inter-op count in compute pool

[DebadityaHait]

Summary

• Treat negative inter_op_parallelism_threads values passed to the process-wide compute pool as unspecified, using the existing default path.
• Add regression coverage that ComputePool creates a valid pool for -1.

Addresses #124954.

Validation

• git diff --check
• CI: bazel test --config=linux -k //tensorflow/core/common_runtime:process_util_test

[gemini-code-assist]

Code Review

This pull request updates InitComputePool to fall back to the default number of threads when inter_op_parallelism_threads is less than or equal to zero, and adds a unit test to verify this behavior. The feedback suggests strengthening the test assertion to ensure the thread pool size matches DefaultNumInterOpThreads() exactly, rather than just verifying it is greater than zero.

[gemini-code-assist]

The assertion EXPECT_GT(pool->NumThreads(), 0) is very weak and only verifies that the pool has at least one thread. Since the goal of passing a negative value is to fall back to the default number of threads, we should explicitly assert that pool->NumThreads() equals DefaultNumInterOpThreads(). This provides a much stronger guarantee that the default path was correctly taken.

Suggested change

	EXPECT_GT(pool->NumThreads(), 0);
	EXPECT_EQ(pool->NumThreads(), DefaultNumInterOpThreads());