[Pallas] Downgrade compact_worklist grouping gracefully when kernel doesn't match

[norx1991]

Summary

pallas_worklist_grouping became an autotune knob independent of pallas_loop_type in #3131. Its config-space gating (offer whenever the kernel has a nested hl.tile shape) is looser than detect_compact_worklist_plan's requirement (must have an hl.grid owner), so autotune can pick pallas_worklist_grouping in (1, 2) for kernels the compact-worklist detector can't handle. In that case detect_compact_worklist_plan raises exc.InvalidConfig("compact_worklist could not locate the owner hl.grid loop in source.") — and that raise escapes past the autotuner's skip path (it's called via the pre-codegen setup, not via a config-compile-check), FAILing the whole sweep step.

This regressed rms_norm-bwd[8192,8192] in the 07-28 TPU sweep (was 2.62× on 07-07, now FAIL). Reproduces on any kernel with the nested hl.tile(x, block_size=…) + hl.tile(mb_cta.begin, mb_cta.end) shape but no hl.grid — e.g. examples/rms_norm.py::rms_norm_bwd.

Catch exc.InvalidConfig in _setup_compact_worklist and leave env.compact_worklist_plan = None. Downstream lowering paths already gate on plan is not None, so grouping silently downgrades to a no-op for kernels that can't use it. Matches the "autotuner-skippable" semantic the raise's docstring already documented.

Verification

• Regression test added (test_grouping_downgrades_on_kernel_without_hl_grid): confirmed FAILs on main with the exact InvalidConfig message from the sweep and PASSes with this fix.
• Pre-existing test updated: test_unsupported_kernel_raises → test_unsupported_kernel_downgrades. It asserted the raise this PR intentionally removes; it now asserts to_triton_code succeeds and emits no compact-worklist builder. test_invalid_worklist_grouping_raises still passes, confirming the try/except doesn't swallow the separate invalid-value validation. Full test_pallas_worklist.py: 86 passed on TPU v7.
• End-to-end on TPU v7 (full autotune + --subprocess-per-shape):

Shape	07-07	07-28 (main)	With fix
rms_norm-bwd[2048,4096]	1.37×	1.60×	1.58×
rms_norm-bwd[8192,8192]	2.62×	FAIL	2.85×