[Pallas] Keep concrete dim sizes concrete in slice/reduction shape inference

[norx1991]

Summary

Stacked on #2279.

When a Pallas kernel does x[tile, :] over a host-origin tensor and then combines the result with a host-allocated buffer of the same trailing extent (e.g. acc = x.new_zeros([n]); acc + sum_result), shape inference used to allocate a fresh unbacked symbol for the slice's trailing dim while the host accumulator stayed concrete (post-#2279). The two sides could not be unified at trace time, raising broadcast errors at the binop:

RuntimeError: Attempting to broadcast a dimension of length 512 at -1!
Mismatching argument at index 1 had torch.Size([512]); but expected shape should be broadcastable to [384]
While processing: grad_w_acc += torch.sum(dy_mb * x_hat, dim=0)

This was previously masked by the dispatch-mode padding (_PadTensorFactoryMode) that forced both sides through next_power_of_2 rounding so they happened to share an unbacked symbol. Disabling that padding for Pallas (#2279) exposes the trace-time mismatch.

This PR:

• _device_indexing_size (type_propagation.py) and SubscriptIndexing.compute_shape (indexing_strategy.py): on backends that don't pad factory ops to power-of-2, keep concrete int slice extents concrete instead of allocating a fresh reduction-dim symbol.
• ReductionLoopBlockSizeSource.from_config (compile_environment.py): route the persistent-reduction extent through backend.static_rdim_size so the host-side _RDIM_SIZE_* constant agrees with what the rest of the codegen computes.

Triton and CuTe paths are unchanged: pad_factory_tensors_to_power_of_2 stays True for them, so they keep allocating reduction-dim symbols and next_power_of_2-rounding the persistent-reduction extent.

Together with #2279, this unblocks tracing of examples/layer_norm.py::layer_norm_bwd on Pallas with non-power-of-2 trailing dims.

Test plan

Added test/test_examples.py::TestExamples::test_layernorm_bwd_non_pow2_dim (Pallas-only, runs in CI). Verified locally that without this PR the test fails with the broadcast error above; with this PR + #2279 it passes. All existing Pallas tests in test/test_pallas.py still pass (88 pass, 4 xfail unchanged).

[norx1991]

Closing for now — the "keep concrete dim sizes concrete" change in this PR breaks test/test_pallas.py::TestPallas::test_nested_fori_loop_scratch_scoping (ShapeMismatch: [u2, 8, 256] vs [u3, 8, 256]). The mechanism is that bypassing allocate_reduction_dimension for concrete-int slice extents perturbs symbol allocation across nested-tile broadcasting in a way I haven't fully understood yet. The u1 shape-prop bug at non-pow2 trailing dims still needs a fix, but a different shape than this. #2282 has been re-stacked directly on #2279 so the per-tile element cap fix can land independently.