feat: fetch multiple Figma nodes in one get_figma_data call

[GLips]

What this enables

get_figma_data can now fetch several top-level nodes in a single call by passing comma-separated ids (nodeId: "1:2,3:4"). Previously, assembling a multi-frame design meant one tool call per node — more round-trips, more latency, more tokens spent re-sending the file context each time. The LLM driving the tool can now grab every frame it needs at once.

Two audiences benefit:

• Tool callers (LLMs / end users): multi-node fetch in one request; and on a partial miss (some ids resolve, some don't) the result now carries a MISSING_NODES marker so the model knows it's working from an incomplete design instead of silently assuming it got everything.
• Developers in this code: parseAPIResponse is no longer a two-branch mutation bag — both API response shapes (full-file and by-nodes) normalize into a common slice list and aggregate declaratively. This PR also lands the FigmaService characterization tests that lock down the request/error path the feature exercises.

Why now

The Figma REST API's nodes?ids= endpoint has always accepted a comma-separated list, but the parser only ever read the first node and dropped the rest. The same line also crashed with a raw TypeError when the API returned an empty nodes object — bypassing the friendly not-found guidance one line below it. This is a captured, grounded request (docs/ideas/multi-node-fetch.md), and the fix for "only reads the first node" is the same edit that removes the crash.

What reviewers should know

• Bundles two advisor plans: the FigmaService request/error characterization tests (the safety net for the request layer) plus the multi-node feature itself. Squash-merges to one commit on main.
• getRawNode is deliberately untouched — it already passes the ids string straight through to the API, so no change was needed at the request layer.
• Cross-node dedup was the main correctness risk and it's verified. Style hoisting and element templating run over the combined forest (one shared globalVars/traversalState), so a style or subtree shared between two separately-requested nodes is counted across both and elevated once — the same machinery that already deduped across a full file's top-level frames. Tests cover the shared-hoist, single-use-inline, and cross-root-template cases.
• Partial-miss handling: unresolved ids are skipped (not fatal) as long as at least one node resolves; only an all-miss or empty response raises the tagged not_found error. The skipped ids are logged server-side and surfaced to the caller.
• Out of scope (intentionally): no change to the output wrapper shape beyond the optional missingNodeIds/MISSING_NODES field, and download_figma_images keeps its own single-node regex (it takes an array, not a comma scalar).

Reviewed for correctness across two independent passes (incl. a cross-model audit); full suite, type-check, and lint green.