Fix workspace allocation and reuse in matmul_all_reduce

[MashrafeeAryan]

Rationale for this change

When one or more GPUs are used, if one GPU needed new buffers while another skipped
allocation, the operation could hang and not proced,

Fixes #464.

What changes are included in this PR?

• Separate buffer allocation from preparation for each calculation.
• Make all GPU workers decide together whether they need new buffers. If any workers need them, all workers allocate memory
• For one-shot and two-shot methods, give each calculation a number. Use that number to track when its results are ready and when all workers have finished using them. This allows buffers to be reused safely without clearing them every time.
• Update benchmarks to let the operation manage preparation.
• Add GPU tests for collective allocation and repeated calculations.

Are these changes tested?

GPU tests and performance measurements have not been run because
no AMD GPUs are available locally.