Export SparseCalibrationWeights and add optional seed parameter

[MaxGhenis]

Summary

• Export SparseCalibrationWeights from the top-level l0 package — it was already documented in the README and in CRITICAL_TEMPERATURE_BUG.md as the temperature-correct path for calibration but was missing from __all__. SparseL0Linear was already exported.
• Add optional seed: int | None kwarg to SparseCalibrationWeights and SparseL0Linear. When set, log_alpha jitter (and log_weight jitter inside SparseCalibrationWeights.fit) draws from a local torch.Generator; seed=None preserves the legacy global-RNG behaviour.
• Update CLAUDE.md to list calibration.py / sparse.py as first-class modules, update test file layout, and replace the stale black -l 79 formatter hint with the current ruff format invocation.

Addresses bug-hunt findings

• Add group sparsity support #5 LOW SparseCalibrationWeights missing from l0/__init__.py::__all__ and CLAUDE.md listing stale module set.
• Implement progressive pruning schedule #6 LOW No seed parameter on SparseCalibrationWeights.__init__ / SparseL0Linear.__init__; jitter relied on the global RNG.

Test plan

• uv run pytest tests -x -q passes (88 passed, 1 skipped; previously 83).
• test_sparse_calibration_weights_exported pins "SparseCalibrationWeights" in l0.__all__.
• test_seed_produces_deterministic_log_alpha and test_seed_none_uses_global_rng verify both RNG paths in test_calibration.py and test_sparse.py.

[MaxGhenis]

LGTM.

Verified:

• l0/__init__.py: SparseCalibrationWeights added to __all__ alongside SparseL0Linear (previously only the latter was exported). Import works: from l0 import SparseCalibrationWeights succeeds; l0.SparseCalibrationWeights is SparseCalibrationWeights.
• seed: int | None = None on both SparseCalibrationWeights.__init__ and SparseL0Linear.__init__, with matching docstring entries.
• Default None preserves legacy global-RNG behaviour — test_seed_none_uses_global_rng pins this with torch.manual_seed(0) on both sides.
• When seed is set, a local torch.Generator is created (on the module's device) and used consistently in both init jitter and (for SparseCalibrationWeights) the fit log_weight jitter. The randn_like-has-no-generator edge case is handled explicitly via torch.randn with shape + generator.
• SparseL0Linear.__init__ also rewrites torch.normal(mu.item(), 0.01, size=...) to mu.item() + 0.01 * torch.randn(..., generator=...) so the seeded path has a single consistent RNG — nice tidy.
• CLAUDE.md now lists calibration.py / sparse.py as first-class modules, updates the package tree, adds test_calibration.py / test_sparse.py, and replaces the stale black -l 79 formatter hint with ruff format. Subsumes finding #5's doc staleness.
• 28 tests pass locally.

Will go green once #44 merges.