Wire max-shape inference into workspace estimation (#31613)
### Description
This PR wires estimation-only input-shape hints into the two-level
workspace-estimation flow described in #29775.
It adds the `session.max_shape_override` session option:
```text
input_ids:[8,4096];attention_mask:[8,4096]
```
ORT validates the configured graph-input shapes before cloning, applies
them to a disposable shadow graph, runs recursive shape inference, and
captures concrete intermediate `NodeArg` shapes without changing runtime
input constraints or executable-graph metadata.
The resolved shapes are consumed at:
- **Level 1:** CUDA `MatMulNBits` workspace estimation during
`GetCapability()`.
- **Level 2:** `DeclareWorkspaceRequirements()` on constructed kernels
during `FinalizeSessionState()`.
Static graph shapes remain usable at Level 2 without configuring an
override. Nodes whose inputs remain unresolved retain the existing
dynamic allocation/estimation fallback.
The shadow workflow:
- Shares existing initializer `OrtValue`s instead of copying large
weights.
- Preserves model-local functions and dynamically generated schemas.
- Captures main-graph and control-flow-subgraph shapes under separate
graph identities.
- Matches source and shadow control-flow nodes by stable identity
instead of topological position.
- Reuses results when the graph is unchanged and rebuilds them after
layout, fusion, or partitioning mutations.
- Converts shadow-model construction and resolution exceptions to
`Status`.
Invalid overrides fail with explicit errors, including unknown inputs,
rank mismatches, static-dimension conflicts, inconsistent symbolic
dimensions, empty entries, and malformed dimension lists. Minimal builds
warn and ignore the option rather than failing session initialization.
The inferred downstream shapes are estimation hints, not guaranteed
upper bounds: operator shape transformations are not necessarily
monotonic. Future preallocation must retain runtime bounds checks and
allocation fallbacks.
Level-2 declarations are currently observational. Workspace offset
planning and preallocation remain follow-up work.
### Motivation and Context
Dynamic transformer models commonly retain symbolic batch and sequence
dimensions, preventing workspace requirements from being calculated
before execution.
Deployment shape hints make partition-time budgeting and post-kernel
workspace declaration possible while preserving dynamic runtime
behavior. The isolated shadow graph prevents optimizers and runtime
validation from treating those hints as the model's static shape
contract.
This is part of the constrained-environment memory roadmap tracked by
#29775 and documented in
`docs/annotated_partitioning/future_directions_constrained_env.md`.
Kernels that do not implement `DeclareWorkspaceRequirements()` retain
the default no-op and continue using their existing scratch-allocation
path.
### Testing
- Built the Debug `onnxruntime_test_all` target on Windows.
- Ran all 20 `MaxShapeOverride.*` tests.
- Added coverage for parsing, malformed separators, zero extents, shape
propagation, canonical-graph isolation, invalid overrides, and stable
matching of parallel control-flow subgraphs.
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 1ceed57f-bf2a-45c7-a865-05bc6c4a598f
Copilot-Session: 3915a111-4266-46e9-9922-3f26d253bbf8
