[Fix] Key compiled artifacts by the sizes they were specialised on

[cennn]

Stacked on #49.

Problem

An artifact is keyed by runtime_shape alone, which is None for every dynamically compiled graph. But the dimensions nobody marked dynamic are not dynamic: _mark_static_shapes burns them into the graph as constants. So a graph traced for one input size is silently handed to another, and the mismatch surfaces far from its cause, as a stride assertion inside the generated kernel:

AssertionError: expected size 75==7, stride 8160==8160 at dim=2
This error most often comes from a incorrect fake (aka meta) kernel for a custom op.
  File ".cache/magi_compiler/magi_cache/model_1_decode_rank_0_cp1_dp1_pp1_tp1/5a9aca216c/artifact_shape_None_subgraph_0/...", line 6779, in call

A VAE decode graph traced for a 1s video clip (7 latent frames) was loaded for a 12s one (75 frames): both resolve to artifact_shape_None_subgraph_0 under the same model tag. The graph signature confirms only two dims generalise — submod_0(..., l_z_, s37, s38, ...) — the declared dynamic_arg_dims={"z": [3, 4]}, while dim 2 is a constant.

Nothing reports a miss; the load succeeds and the failure looks like a bad meta kernel.

Fix

CacheEntry and the artifact path now carry a digest of the sizes that were specialised. Traces differing only in those sizes stop colliding: the second one misses, compiles, and stores under its own key. Dimensions marked dynamic arrive as SymInt and are excluded from the digest, so genuinely shape-agnostic artifacts are reused exactly as before, and no extra compilation is introduced — dynamo already retraces on those shape changes, this only stops the disk cache from serving the wrong result across processes.

Verification

1s : 89e2381f22
12s: f8a6bb7f7a
1s again: 89e2381f22

End to end on athena's gaga4 pipeline: the 12s clip that reproduced the assertion above now decodes through, 8 requests at 73s each.

Same class of defect as #42 (parallel topology into the cache directory) and #49 (compile state per topology): a key that omits something the artifact actually depends on.

[cennn]

Closing: this targets the wrong layer.

compile_sizes defaults to an empty list and nothing sets it (athena has no occurrence; the only reference in this repo is a test passing []), so concrete_size_entries is always empty and PiecewiseBackend.call always falls through to compiled_graph_for_general_shape. There is only ever one graph in memory, so a cache key that distinguishes shapes only helps on the narrow reload-after-restart path.

Measured: with this change already applied, a VAE window traced at 8 latent frames was still handed a 9-frame input in the same process and died on the same kind of assertion (expected size 9==8 at dim=2). The key did not prevent it.

The real constraint is that a magi_compile'd callable admits exactly one combination of specialised sizes, because _mark_static_shapes burns every undeclared dimension into the graph while the backend treats the first trace as general. Missing dynamic_arg_dims should fail loudly and name the dimension, not be papered over by a key. The crash that motivated this was fixed on the athena side by declaring the time dimension dynamic (world-sim-dev/athena PR #955).