[release/2.7][SWDEV-544125] update test buffer fudge factor for hipblaslt for test_fully_shard_training_memory test (#2493)

In this PR, I cherry picked upstream commit
78300c8.
This fixes the test_fully_shard_training_memory test under
/distributed/_composable/fsdp/test_fully_shard_memory.py.
It was a failing test in Jira
https://ontrack-internal.amd.com/browse/SWDEV-544125

Co-authored-by: Ethan Wee <[email protected]>

Author: akashveramd

test/distributed/_composable/fsdp/test_fully_shard_memory.py

@@ -117,6 +117,9 @@ def _test_fully_shard_training_memory(
         # number is kept much smaller than the actual memory usage, which is on
         # the order of 100-200+ MB)
         buffer_mb = 16
+        # The default workspace for hipblaslt is larger than for cublas/cublaslt
+        # which requires a slight increase to this buffer value.
+        buffer_mb = 16 if torch.version.cuda else 18
         if reshard_after_forward:
             # 3x max unsharded block parameters (current all-gather + copy-out
             # and next all-gather), non-block parameters, and other