fix(tests): make -k <registry-key> work in smoke test

[ahsi5086]

Problem

pytest -k gemma4_moe collected 0 tests despite gemma4_moe being a
valid registry entry. The smoke test was parametrized with raw HF paths
as test IDs (e.g. google/gemma-4-26B-A4B-it), so -k gemma4_moe had
nothing to match against. Only models whose HF path happened to contain
the key name as a substring (e.g. qwen3 → Qwen/Qwen3-0.6B) worked
by accident.

Fix

Introduce CAUSAL_PARAMS — a list of pytest.param(path, id=key)
objects where the registry key name is the test ID. Switch
test_e2e_smoke_spyre.py to use CAUSAL_PARAMS instead of
CAUSAL_PATHS. Update xfail_non_blocking to handle both plain strings
(all existing callers are unchanged) and pytest.param inputs.

CAUSAL_PATHS is untouched — every other test file that uses
ids=CAUSAL_PATHS continues to work exactly as before.

Add scripts/run_smoke_by_key.py for models intentionally excluded from
the default collection by the smallest-per-adapter selection rule (e.g.
granite8b loses to granite2b for hf_granite.py). These can't be
reached via -k since they're never collected — the script runs the
smoke test for any registry key directly.

Usage

# Now works — previously collected 0 tests
pytest -s -vvv tests/spyre/test_e2e_smoke_spyre.py -k gemma4_moe

# Also works for any other registry key
pytest -s -vvv tests/spyre/test_e2e_smoke_spyre.py -k qwen3
pytest -s -vvv tests/spyre/test_e2e_smoke_spyre.py -k granite2b

# For models not in the default collection (shares adapter with smaller model)
python scripts/run_smoke_by_key.py granite8b
python scripts/run_smoke_by_key.py --list   # show all available keys

[arielge]

@ahsi5086 actually the most consistent approach IMO is just to fix the documentation (#451), the entire testing mechanism relies on paths