feat: bound class sampler - #253

A class's run was picked uniformly among its runs, then a start uniformly inside
it, so every run contributed one equally likely position regardless of size.
With a 400-row run beside a 40-row one, each got half the chunks and the short
run's rows were drawn 10x as often per row. Class marginals were unaffected, so
no share test noticed.

The ClassSampler docstring still describes the intended scheme: "a uniform
chunk-start within c is drawn per chunk (a prefix-sum lookup maps it to the
absolute slice in O(log n_runs))". The sampler worked that way until ec49f9b,
where the prefix sum changed from cumulative valid positions to cumulative run
counts, leaving searchsorted nothing to resolve against. It was replaced by a
direct index plus a second draw.

Each run now records how many chunk starts it holds and how many precede it in
the table, and one uniform draw per class resolves through those to a run and an
offset inside it. Two rng.integers calls become one, so a given seed yields
different slices than before.

Both class samplers draw through RLEManager.slices_from_classes, so this fixes
WeightedClassSampler too, and the new tests run over both.

Replays another class sampler's per-batch class schedule against a second
annotation column, so one pass reads two annotations in step with each other.
It takes a ClassSampler or another BoundClassSampler -- so these chain -- and
rejects WeightedClassSampler, whose batches mix classes by design.

The shared machinery moves into two private bases in _class_sampler.py:
_RunClassSampler holds the constructor, the remainder handling and the window
loop, and _ScheduledClassSampler adds the per-batch schedule that makes a
sampler replayable. WeightedClassSampler drops its near-duplicate
_iter_requests and overrides only the two policy hooks, which also removes the
Liskov problem of it subclassing ClassSampler while batching differently.

BoundClassSampler keeps a private copy of its inner sampler, so nothing the
caller does to their own instance afterwards can desynchronise the schedule;
its rng setter forwards to that copy, so re-assignment -- as DistributedSampler
does per rank -- still reaches it.