Preflight disk headroom per node; raise cleanly when it won't fit

[bscholer]

Summary

Closes the two failure modes the periodic auto-sweep (#178) couldn't address on its own:

1. cache_max_bytes set to the whole disk. The 30s sweep finds nothing to evict (cache is under budget), but the next op still ENOSPCs because actual `disk_free` is way smaller than the budget headroom suggests. Veil Nebula at 1.5TB was this case.
2. The op genuinely needs more bytes than we can possibly free. Should surface a clean "needs N GB, only M GB available" error before the job starts, not a half-stacked failure deep into Siril.

Shape

• `Node.estimate_storage_bytes(inputs, params)` is a new opt-in method on the base class (`None` by default). Bulk-tier sequence nodes (`calibrate`, `convert_lights`, `seq_bg_extract`, `seq_offset`, `seq_register`, `seq_resample`) implement it using Siril's own size formula from `src/io/sequence.c`: `frame_bytes = rxrynb_layers*bps + FITS_HDU; total = frame_bytes * nb_frames`. We approximate by scaling the input directory's on-disk bytes for bit-depth (uint16→float32 is 2x) and spatial scale (drizzle 2x is 4x, resample 0.5 is 0.25x). One sample FITS header read tells us BITPIX.

• `ensure_disk_headroom(cache, need_bytes, cache_max_bytes, db_path)` enforces `headroom = min(disk_free, cache_max_bytes - cache_used) >= need`. Both constraints are binding — whole-disk-cache case falls out naturally because `cache_max - used` becomes huge and `disk_free` binds.

• Runtime calls the estimator right before `reserve()`, sweeps aggressively if needed, emits `node_failed` with a clean message on raise. Also logs `estimate=X actual=Y ratio=R` on every successful bulk node so you can validate the multipliers on real datasets and tune from data rather than guessing.

• `JobWorker` passes `cache_max_bytes` + `db_path` into `run_job` so the headroom path activates. Tests that bypass JobWorker get the legacy no-op behavior.

Drizzle is handled

`seq_register` passes `scale=drizzle_scale` (1, 2, or 3) when drizzle is on. A drizzle-2 register against a 100GB input estimates 400GB (×4 area), so the preflight catches that before it ENOSPCs mid-job.

Test plan

• 10 new tests in `tests/test_storage_headroom.py` covering estimator (frame count, bit-depth, scale), headroom helper (noop / sweep / raise / budget-binds vs disk-binds)
• Full suite: 678 passed, 1 skipped
• Ruff clean
• After deploy: run a real Veil Nebula register on the Linux box, watch the logs for the estimate vs actual ratio. Tune the bit-depth multiplier if the ratio is consistently off.