Fix pass/input indexing and remove model-only execution assumptions

Problem

SGProcessingManager currently mixes input indexes with pass indexes and assumes every pass has a model.

Confirmed examples:

• GetCidForProc() resolves an input index with GetInputIndex() and then uses that index to access processing_.get_passes()[index].
• ParseBlockSize() calls pass.get_model().value() for every pass, even though COMPUTE, DATA_TRANSFORM, and RENDER passes are allowed without a model.
• Several optional fields are accessed through .value() without pass-specific validation.

These assumptions can select the wrong pass, access an out-of-range pass, or fail on valid non-inference jobs.

Scope

• Give passes stable indexes or IDs independent of global input indexes.
• Resolve inputs through each pass's declared bindings.
• Remove unconditional model access from generic processing paths.
• Make block-size and subdivision logic depend on the active pass and its bound resources.
• Add bounds checks and useful errors for missing pass, input, output, model, shader, dimensions, and source fields.
• Preserve current MNN inference behavior while making non-model pass types safe.

Acceptance criteria

• Jobs with different numbers of passes and global inputs resolve the correct pass and resources.
• COMPUTE, DATA_TRANSFORM, and RENDER definitions do not fail merely because they lack a model.
• Invalid references return typed errors rather than throwing or reading the wrong array element.
• Tests cover one-to-many and many-to-one input/pass layouts, missing optionals, and out-of-range references.

Base implementation and review work on the branch consumed by SuperGenius/develop.