Add unit tests for xrex/models/loss_recsys.py

[AnonRish]

Summary

• Add 17 tests covering multihot_loss_compute, continuous_loss_compute,
  and tweedie_loss_compute
• Add a dev dependency group (pytest) to pyproject.toml, since nothing
  in the project currently declares a test runner

Problem

phoenix/xrex has 165 Python files and, as far as I could find, zero test
coverage (no tests/ directory anywhere under phoenix/, and grepping for
pytest/unittest usage turns up one incidental hit). loss_recsys.py --
the multihot, continuous (mse/mae/huber), and Tweedie loss functions used
in training -- is pure and deterministic, so it's a good place to start.

Coverage

Correctness on hand-checkable inputs (perfect predictions -> near-zero
loss); masking behavior (padding, negative-sample exclusion); weighting;
clamping and flooring of inputs; all three loss_type branches of
continuous_loss_compute; all three p-value branches of the Tweedie
deviance, checked against the closed-form Poisson and Gamma special cases
at p=1 and p=2; error handling for an unknown loss_type.

Verification

uv sync --group dev
uv run pytest xrex/models/test_loss_recsys.py -v
# 17 passed

Notes for reviewers

multihot_loss_compute's tests need an explicit single-device JAX mesh and
NamedSharding-wrapped inputs (see the _shard() helper) because of its
internal with_sharding_constraint call -- this isn't test scaffolding
overhead added for style, it's what the function actually requires even
for a single-array, single-device call. Flagging in case that's not
intentional/known. Separately, raw_weights in the same function must be
shaped like padding_mask ((batch, seq)), not a flat per-row array -- a
shape mismatch broadcasts silently into something numerically meaningless
instead of raising. Not fixed here (out of scope for a test-only PR);
worth a one-line docstring note if someone sends a follow-up.