Avoid lazy init of Blas handles, fix for non-canonical dots

[pemeliya]

This is a backporting of these 3 upstream PRs:

• Avoid lazy init of Blas handles: [ROCM] Avoiding lazy initialization of blas handles on ROCM openxla/xla#21616
• Fix typo in buffer_sharing: [ROCM] Bugfixing typo in buffer_sharing openxla/xla#21620
• Fix for non-canonical dots: [ROCM] Fixing non-canonical dots and enabled conv_add_multiply_reorder on ROCM openxla/xla#21618

Partitularly, this shall fix the issue in: https://ontrack-internal.amd.com/browse/SWDEV-482895

[i-chaochen]

For https://ontrack-internal.amd.com/browse/SWDEV-482895 we need to backport to 0.4.30

[pemeliya]

Sounds like we have some serious RCCL problem on CI nodes. Several multiGPU tests seem to fail with (e.g. functional_hlo_runner_test):

19:21:31  E0000 00:00:1737397209.417337   83920 pjrt_stream_executor_client.cc:3085] Execution of replica 0 failed: INTERNAL: xla/service/gpu/runtime/nccl_api.cc:515: NCCL operation ncclGroupEnd() failed: internal error - please report this issue to the NCCL developers. Last NCCL warning(error) log entry (may be unrelated) 'Failed to find reverse path from remNode 0/11000 nlinks 2 to node 0/ae000'.
19:21:31  E0000 00:00:1737397209.417412   83923 pjrt_stream_executor_client.cc:3085] Execution of replica 0 failed: INTERNAL: xla/service/gpu/runtime/nccl_api.cc:515: NCCL operation ncclGroupEnd() failed: internal error - please report this issue to the NCCL developers. Last NCCL warning(error) log entry (may be unrelated) 'Failed to find reverse path from remNode 0/11000 nlinks 2 to node 0/ae000'.

[i-chaochen]

I think @hsharsha mentioned that is because few nodes are broken....