Unified worker sizing, resilient bulk OSM/NLD downloads, faster branch processing

[supathdhitalGEO]

Parallelization

• New fimbox._workers: one place that answers "how many workers can this machine
  actually run?" Auto-sizes to min(cpu, RAM / per_worker), honours an explicit 1
  for serial debugging, clamps optimistic requests, and never spawns more workers than
  tasks. Overridable via FIMBOX_WORKERS / FIMBOX_RAM_PER_WORKER.
• n_workers now defaults to None ("size to the machine") across branch processing,
  FIM generation, and every rating-curve calibration step — previously hardcoded 1/4.
• _dask.py sizing now delegates to the shared resolver, so n_workers=64 on a laptop
  lands on what the laptop can feed instead of thrashing swap.
• Bridge LiDAR fetching shares a single bounded thread pool (n_workers × tile_workers,
  capped) with a tile cache, instead of a pool per footprint.

Bulk OSM download

• Shared _OverpassClient for roads and bridges: probes all mirrors in parallel up
  front, permanently drops dead hosts for the session, and treats an Overpass remark
  (query timed out inside an HTTP 200) as a failure rather than partial success.
• AOI tiling with off-polygon tiles dropped before any request, round-robin across live
  mirrors, and quadrant sub-splitting for tiles that fail — this is what makes large
  AOIs work.
• Raises instead of writing a file that looks complete but has holes (allow_partial=True
  to override).
• Bridges now come straight from Overpass rather than osmnx, with a curated tag schema
  (no list columns / duplicate-cased columns that used to break GPKG writes).

NLD download

• Layer ids resolved by name at runtime (fallback to 15/16), so a USACE renumbering
  warns instead of silently returning nothing.
• New NLDQueryError keeps "failed download" distinct from "this AOI has no levees" —
  a missing levee burn quietly changes the HAND surface. Layers are attempted
  independently and failures reported loudly.
• Levee Z-geometry path now reuses the paged, shrink-on-failure fetch.

Performance

• Numba-JIT D8 routing shared via _d8.downstream_index, replacing deque-based
  traversal in flow accumulation and gage-catchment labelling.
• SRC filtering precomputes reach ladders once instead of per-key.