feat(heads): add multitask loss and sampling primitives

[jsugg]

Groundwork for masked multitask training and balanced batch construction. Both
modules are deliberately unwired — nothing in product code imports them yet,
so this carries no behavioural risk and can be reviewed purely on its own terms.

MaskedUncertaintyWeightedLoss (_internal/heads/multitask_loss.py)

Combines per-sample task losses using only the targets a sample actually marks
available, weighted by learned log variances. Tasks with no available targets in
a batch are skipped rather than contributing zero. The primary emotion task gets
a weight floor (default 0.25) so auxiliary objectives cannot suppress it, and
supplying no available targets at all raises rather than silently returning zero.

Utterance sampling (_internal/models/utterance_sampling.py)

• utterance_sampling_distribution — square-root corpus mass and
  inverse-square-root class mass within corpus, so large corpora and frequent
  classes stop dominating. Probabilities are asserted to sum to one.
• select_training_windows — bounded seeded window subset per epoch, keyed by
  seed:epoch:sample_id, so selection is deterministic and replayable.
• sampling_contributions — reports expected sample and duration contribution
  per corpus and class, for auditing a recipe before training on it.

Why land it unwired

These are reusable research primitives that the Torch trainer will need. Landing
them now keeps them reviewed, typed, and tested instead of drifting in a working
tree. They are covered by unit tests and can be deleted cheaply if the design
changes.

Verification

make lint, make type, and make test pass. Pre-push gate passed.