Support observation windows shorter than the right-truncation PMF

[developer-rpai]

Closes #714.

What this does

compute_prop_already_reported no longer requires the observation window to
cover the full reporting-delay PMF support. When the prediction vector is
shorter than the delay support (minus right_truncation_offset), the function
now returns the trailing slice of the reported-proportion tail, i.e. the
proportions for the most recent timepoints. _apply_right_truncation drops the
ValueError that forced callers to pad short windows, so short observation
windows just work.

Root cause

The function built its output as [ones(n_pad), tail] with
n_pad = n_timepoints - len(tail). For a short window n_pad goes negative,
which died inside JAX with a cryptic TypeError: broadcast_in_dim shape must have every element be nonnegative, and the model-level caller raised its own
ValueError before even getting there. The tail is ordered oldest to newest
with tail[0] = 1.0, so a short window covering only recent timepoints is
exactly tail[len(tail) - n_timepoints:]. Full-window behavior is unchanged.

Reproduction (pristine code)

from pyrenew.convolve import compute_prop_already_reported
import jax.numpy as jnp
compute_prop_already_reported(jnp.array([0.5, 0.3, 0.2]), 2, 0)
# TypeError: broadcast_in_dim shape must have every element be nonnegative, got (-1,)

After the fix this returns [0.8, 0.5], matching the last two entries of the
full-window result [1.0, 1.0, 1.0, 0.8, 0.5].

Tests

• test/test_convolve.py: new test_compute_prop_already_reported_short_window
  (hand-calculated values, including the boundary where the window exactly
  matches the tail length) and
  test_compute_prop_already_reported_short_window_matches_full (short-window
  output equals the trailing entries of the full-window output, over window
  lengths 1..5 and offsets 0..1).
• test/test_observation_counts.py: TestRightTruncation:: test_short_observation_window_raises rewritten as
  test_short_observation_window_supported, asserting via the
  prop_already_reported deterministic site that a 2-day window with PMF
  [0.2, 0.3, 0.5] yields [0.5, 0.2].

Results: test_convolve.py 40 passed, test_observation_counts.py 96 passed
(136 total). ruff check and ruff format --check clean on all touched files.

Note: this sandbox runs Python 3.12 while the package declares >=3.13, so the
editable install was skipped and tests ran with PYTHONPATH instead. The
touched code paths are version agnostic, but CI is the final judge.