Refactor 09_gemm_one_shot_all_reduce example with pytest and testable function

[Copilot]

This PR implements a comprehensive pytest test suite for the 09_gemm_one_shot_all_reduce example and refactors the code to enable proper functional testing in CI environments.

Key Changes

Refactored Example Structure:

• Extracted gemm_one_shot_all_reduce() function from benchmark.py that takes input matrices and algorithm parameters
• Function encapsulates the core GEMM + all-reduce computation logic and returns the result matrix
• Enables reuse between the command-line example (with arg-parsed parameters) and test suite (with parametrized inputs)
• Maintains all original functionality while improving testability

Comprehensive Test Suite:

• Removed error catching for missing AMD GPU libraries since ROCm is expected in CI
• Added functional tests that validate actual GEMM computation results using the extracted function
• Parametrized tests covering multiple data types (fp16, fp32, bf16) and matrix configurations
• Tests validate algorithm requirements like matrix dimension divisibility constraints
• Uses existing validation functions to verify mathematical correctness
• Updated tolerance in tests to match example (atol=2)

Testing Features:

• Tests core algorithm functionality rather than just import capabilities
• Validates block size calculations and parameter handling
• Checks file structure and content organization
• Follows established patterns from other example tests

The refactored code enables robust continuous integration testing of the GEMM one-shot all-reduce algorithm while maintaining the original example's command-line interface and functionality.

Fixes #63.

---

💬 Share your feedback on Copilot coding agent for the chance to win a $200 gift card! Click here to start the survey.

[mawad-amd]

@copilot increase the tolerance and match the one used in the example.

[Copilot]

Updated the tolerance to match the example (atol=2). See commit f3a224d.