[tracking][shortfin] Shortfin reliability / iree integration issues

[renxida]

List of problems I'm encountering:
These are encountered on 0120. They might have been problems for a while, but are recently encountered as I add more shortfin test cases.

• Shortfin invocation on gfx 942 and 90a (not CPU) fails and hangs indefinitely instead of shutting down with ValueError: shortfin_iree-src/runtime/src/iree/base/allocator.c:21: INVALID_ARGUMENT; allocator has no control routine: https://github.com/nod-ai/shark-ai/actions/runs/13120212884/job/36604297357?pr=890. Strangely, the sglang integration tests don't fail this way. I need to look deeper into this.

This one is encountered on iree 0124:

• shortfin fails to initialize with ValueError: shortfin_iree-src/runtime/src/iree/task/executor.c:65: UNIMPLEMENTED; threadless donate-only executor mode not yet implemented (conversation here: https://xilinx.slack.com/archives/C077FDT0002/p1738187925783399)