Fix temperature scaling in HardConcrete deterministic gates

[MaxGhenis]

Summary

• HardConcrete._deterministic_gates was computing sigmoid(qz_logits) without the / temperature scaling, so .eval() output silently ignored the temperature parameter. For PolicyEngine's default temperature=0.25 this was a 4x distortion in log-odds space and broke train/eval consistency for every L0Linear / L0Conv2d / L0DepthwiseConv2d / L0Gate caller, and TemperatureScheduler updates were dropped at eval time.
• The fix uses sigmoid(qz_logits / temperature), matching Louizos et al. 2017 Eq. 11 and the temperature-correct implementations in SparseCalibrationWeights (calibration.py) and SparseL0Linear (sparse.py).
• Adds three regression tests.

Addresses bug-hunt findings

• Initial L0 regularization implementation #1 CRITICAL Temperature dropped in _deterministic_gates (fixed in l0/distributions.py:132).
• Add GPU memory optimization with gradient checkpointing #2 HIGH Tests did not catch the temperature bug (test_deterministic_gates_respect_temperature, confirmed to FAIL on pre-fix code and PASS after).
• Add sparse tensor support for pruned models #3 MEDIUM get_active_prob / get_sparsity used temperature but _deterministic_gates did not (test_sparsity_stats_match_eval_activation pins the consistency after the fix).

Test plan

• uv run pytest tests -x -q passes (86 passed, 1 skipped; previously 83).
• Manually reverted l0/distributions.py and re-ran test_deterministic_gates_respect_temperature — fails with AssertionError: Deterministic gates should depend on temperature.
• test_deterministic_gates_match_reference_formula pins the closed form against SparseCalibrationWeights / SparseL0Linear.
• test_sparsity_stats_match_eval_activation verifies get_sparsity_stats agrees with actual eval-time activation rate.

[MaxGhenis]

LGTM — the critical fix is exactly right.

Verified:

• l0/distributions.py:137 changes torch.sigmoid(self.qz_logits) to torch.sigmoid(self.qz_logits / self.temperature) — matches Louizos et al. 2017 Eq. 11 and the standalone SparseCalibrationWeights/SparseL0Linear closed form.
• Only the _deterministic_gates mean is changed; _sample_gates, get_penalty, get_active_prob were already temperature-aware and are untouched (no over-reach).
• Three regression tests are load-bearing:
  • test_deterministic_gates_respect_temperature — confirmed to FAIL on the pre-fix impl: I monkeypatched _deterministic_gates back to sigmoid(qz_logits) and observed torch.allclose(high_gates, low_gates) == True, which violates the test's not torch.allclose assertion.
  • test_deterministic_gates_match_reference_formula — pins the closed form (the fix's Eq. 11).
  • test_sparsity_stats_match_eval_activation — subsumes finding #3 (reported vs actual sparsity agreement within 0.15).
• Findings #1 (critical), #2 (test gap), #3 (within-class inconsistency) all addressed.
• No unrelated changes; just the 5-line code fix + 89-line test block + changelog.

Will go green once #44 merges and lint unblocks.