PT slice_update in place

[polycrit]

The torch backend implements slice_update with scatter_, so the frame is written into the target's buffer.

Before, the torch backend used the generic fallback, a where over the whole axis. The bounded KV cache of CausalSelfAttention calls slice_update once per decoding step, so every step copied the whole cache.

start is a scalar or has other dims of the target (one position per batch entry / hypothesis, as the bounded KV cache passes it). A start beyond the axis fails (CPU error, CUDA device-side assert, also under graph capture, without host sync), where the generic fallback silently writes nothing.

Where autograd records the write (grad enabled and target or value requiring grad), it uses scatter and writes into a copy. The index stays on the target's device, so there is no host sync.

Tests:

• test_slice_update_torch_in_place: the result shares the target's storage, the value lands at start and the rest stays unchanged.
• test_slice_update_torch_autograd: with a target requiring grad, the target keeps its values and the gradient reaches it.
• test_slice_update_torch_start_per_entry: per-entry and scalar start, the last position, a start beyond the axis fails.
• test_causal_self_att_bounded_cache_overflow_fails: the bounded KV cache writes its last free slot, a step beyond fails.

Supersedes #2025.

[albertz]

Pushed be10e2d, merging #2025 into this PR (#2025 gets closed as duplicate):

• start may have other dims of the target, i.e. one position per batch entry / hypothesis.
  The bounded KV cache of CausalSelfAttention passes exactly that (per-hypothesis lengths),
  and the index_copy_ with reshape(1) failed there (shape '[1]' is invalid for input of size 2).
• scatter_ instead of index_copy_, still in place; scatter where autograd records the write.
• A start beyond the axis fails (CPU error, CUDA device-side assert, also under graph capture, without host sync),
  where the generic fallback silently writes nothing.
• rf.slice_update docstring: start scalar or with other dims of target.

Tests added: test_slice_update_torch_start_per_entry (per-entry and scalar start, last position, beyond the axis fails),
test_causal_self_att_bounded_cache_overflow_fails (last free slot written, a step beyond fails).
At RZ (torch 2.7, CPU), test_rf_array.py + test_rf_attention.py -k "slice_update or att": 33 passed, 1 skipped (CUDA only).
Negative control (only the tests on 522c28c): both new tests fail.