Feature/mvp inference

[dougbrn]

Change Description

MVP of the initial design: column spec + inference only. Training is deferred to a follow-up. Design: docs/design/initial-design.md (design PR).

Includes docs/design/mvp-design.md, a trimmed design covering only what this PR ships; the full design (including training) is on the design/initial-design branch.

Solution Description

• Adds lsdb>=0.11 and torch>=2.4 as core dependencies.
• ColumnSpec / Flat / Nested: maps numeric flat and nested catalog columns to tensors. resolve() validates against catalog meta (nested sub-column types via NestedDtype.column_dtypes) without computing data, reporting every problem at once. This replaces the hand-built column preview from hyrax#1062.
• Nested columns are zero-padded to the batch maximum with <key>_mask and <key>_lengths, vectorized, with optional max_length and head/tail truncation. Empty or null nested rows have length 0. Missing values raise.
• predict(catalog, model, spec, outputs) / ModelDeployment: a Dask WorkerPlugin builds the model once per worker from a ModelSpec (factory + state-dict path, or an in-memory module), in eval() mode on a device resolved on the worker (default "cpu", "auto" opt-in). Returns a lazy lsdb.Catalog with correct meta. Requires a dask.distributed Client.
• Alignment by construction: each partition is processed in order with task-local state only (no shared mutable dataset), and predictions are attached by position, never joined on _healpix_29. Row counts are checked per batch.
• Example: docs/examples/predict_on_catalog.py.

Tests: spec validation (including a no-compute guard), collation against a per-row reference, alignment after a filter that empties partitions plus reordering, duplicate _healpix_29 values, one model build per worker, meta dtypes matching computed data, and failure modes. A deliberately reversed batch order makes the alignment tests fail.

Deferred from the fuller branches: transforms, pad_to/pad_value, null filling, dict outputs, no-client mode, and all of training.

Note: CI only triggers on main, so this PR gets no CI until dev is added to the workflow triggers (design doc §12).

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 95.94017% with 19 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.99%. Comparing base (c9ed114) to head (e391572).

Files with missing lines	Patch %	Lines
src/lsdb_ml/spec.py	95.78%	8 Missing ⚠️
src/lsdb_ml/inference/plugin.py	92.18%	5 Missing ⚠️
src/lsdb_ml/device.py	86.36%	3 Missing ⚠️
src/lsdb_ml/inference/outputs.py	96.07%	2 Missing ⚠️
src/lsdb_ml/collate.py	97.77%	1 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##               dev       #9      +/-   ##
===========================================
- Coverage   100.00%   95.99%   -4.01%     
===========================================
  Files            2       10       +8     
  Lines            7      474     +467     
===========================================
+ Hits             7      455     +448     
- Misses           0       19      +19     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[dougbrn]

Padding here seems questionable, it's not padded globally, but per batch.