Various issues

[andfau-amd]

• CI has been failing (red) for some time now
  • Prevent duplicate convbench configs clobbering eachother when compiled #66
  • Fix TK convbench to work with new iree-turbine API #68
  • Fix TK gemmbench to work with new iree-turbine API #70
  • [Codegen] a 32000x4096x1 MatVec fails to compile iree-org/iree#20446
    • Temporarily disable gemmbench configurations affected by IREE bug #73
  • Fix TK gemmbench after renaming of tkw.reduction() to tkw.iterate() #77
• CI has several kinds of silent failures (tests failing but reported as green anyway)
  • Set nonzero exit status when convbench compilation/execution fails #64
  • Set nonzero exit status when gemmbench compilation/execution fails #71
• gemmbench: Lack of FP8 types
  • gemmbench: Generate benchmarks from supplied dtypes instead of filtering #76
  • gemmbench: Add support for four FP8 types and test one in CI #78
• gemmbench: Lack of dynamic shapes
  • gemmbench: Add support for dynamic dims, use for llama8b_prefill #80
• gemmbench: Lack of runtime correctness checks
  • Idea would be to use numpy to run part of the GEMM (maybe just for a single cell) and check against IREE output. Probably not possible with iree-run-benchmark because it doesn't return the output.
• Support for targeting different GPUs (and appropriate filtering/modification of tests for this)
  • This is essentially solved for gemmbench, it already has target and device options, and gemmbench: Generate benchmarks from supplied dtypes instead of filtering #76 limits the datatypes
  • For convbench, needs [CONV] Pass the device and target info from the command line. #24 at a minimum
• Code hasn't been run through automated formatting (diffs in several files)
  • Reformat all files with black and add a README reminder #62
  • Add pre-commit hooks to do reformatting automatically #63
• One or more source code files use CRLF instead of LF for some reason
  • Replace CRLF with LF in wave_conv_utils.py #69