feat(nodes): Add a LoRA collection picker node that selects and applies an indefinite number of LoRAs

[lstein]

Summary

Using LoRAs in workflows currently requires a fixed number of Apply LoRA nodes, and every node must have a LoRA selected or the graph won't fire — making variable-LoRA workflows awkward. This PR extends the Apply LoRA Collection - * nodes to allow users to select and apply any number of LoRAs, or none at all.

• A model picker adds a LoRA, filtered to the node's compatible base model(s).
• Each row has a weight input and a trash icon to remove it.
• The list can be empty and still valid, so the node fires with zero LoRAs.

Existing connections still work (Select LoRA → collection loader, Collect → collection loader), because the LoRAField type name is unchanged — the field simply became stateful/editable in addition to being connectable.

How it works

LoRAField is now a stateful field type whose value is a list of { lora, weight }. The collection loaders' loras input gains ui_model_base/ui_model_type metadata, which the new inline editor reads to filter the picker. The frontend reads ui_model_base from the live OpenAPI schema, so no schema.ts regen is needed.

Backend

• Added ui_model_base + ui_model_type=LoRA to the loras field of every *LoRACollectionLoader (SD1.5, SDXL, FLUX, Anima, Z-Image, Qwen Image, Flux2 Klein); bumped node versions.
• Fixed a pre-existing base-validation bug: the FLUX and FLUX.2 Klein collection loaders both carried an identical copy-pasted assert lora.lora.base in (Flux, Flux2), so each accepted the other architecture's LoRAs even though FLUX.1 and FLUX.2 Klein are separate pipelines with non-interchangeable LoRAs. Each now validates against its own base only, via an explicit ValueError (helpful message, not stripped under python -O) matching the Anima/Z-Image loaders.

Frontend

• Registered the stateful LoRAField type (zod schemas, guards, unions) in field.ts.
• Collection template builder + empty-list default; fieldLoRACollectionValueChanged reducer.
• New LoRAFieldCollectionInputComponent (reuses ModelPicker, useLoRAModels, zModelIdentifierField.parse, DEFAULT_LORA_WEIGHT_CONFIG); wired into InputFieldRenderer.
• Added unit tests for the new schemas/guards.

QA / Testing

• Backend: ruff check + ruff format --check pass on all changed files.
• Frontend: pnpm lint (tsc / eslint / prettier / dpdm / knip) all pass; 1144 unit tests pass, no type errors.
• Manual: confirmed the inline +/weight/trash editor renders on the collection nodes, the picker is filtered per base (FLUX shows only Flux, not Flux.2 Klein), and an empty collection still lets the graph run.

Notes for reviewers

• Reload the app after deploying so the frontend re-fetches the updated OpenAPI schema (the picker filter reads ui_model_base at runtime).