[FlyDSL] Remove obsolete availability helpers

[coderfeli]

Summary

• Remove is_flydsl_available and the duplicated aiter.ops.flydsl.utils layer now that FlyDSL is a required dependency.
• Read LDS limits from FlyDSL's SMEM_CAPACITY_MAP and remove dead buffer helper exports.
• Move tuned CSVs, the MXFP4 v2 tuner helper, and FlyDSL tests into their canonical config, tuner, and test directories.
• Standardize the relocated GDR decode, A4W4 MoE, and split-K HGEMM tests with @benchmark, run_perftest, fp32 checkAllclose, TFLOPS/TB/s metrics, argument sweeps, and Markdown summaries.

This is a dependent PR targeting pr-4789 because the cleanup applies to files introduced or changed on that branch.

Test plan

• python3 -m compileall -q aiter op_tests csrc
• Ruff and Black checks for all changed Python files
• pytest -q op_tests/tuning_tests/test_asm_splitk_guard.py op_tests/test_gemm_a8w8_bpreshuffle_pad_k.py (10 passed)
• pytest -q op_tests/flydsl_tests/test_flydsl_linear_attention.py (16 passed)
• FlyDSL GPU compile/run sweeps for GDR decode, split-K HGEMM, and all four A4W4 MoE paths (err=0)
• FlyDSL GPU smoke tests for causal conv1d and silu/quant paths
• Start all three affected tuning CLIs with --help
• Verify a gfx950 row loads from the relocated GDR tuned CSV

[coderfeli]

Superseded by #5116, rebuilt directly on top of main.