Hold Diffusers pipelines in the engine's local object store

[cjkindel]

Stacked on #66, which moves Diffusers into a worker-v2 execution environment but keeps the library's own
ModelCache. This PR is the second half: built pipelines move out of that cache and into the engine's
local object store.

The object cache

The library kept its own module-global ModelCache. Built pipelines now live in the worker's local
object store through node.local_objects:

• Releasing GPU memory became the engine's job. on_drop=clear_diffusion_pipeline runs on
  displacement, node delete, and worker teardown. Before, release happened only if a node reached
  remove_pipeline on the way past.
• One pipeline stays resident, as before. Building releases the one already held, which is what
  ModelCache.get_or_build_pipeline did when it cleared the whole cache on a new config hash. Holding two
  models' worth of VRAM would be a different experience, not just a new environment. The derived-pipeline
  path keeps its base rather than cascading, since a derived pipeline shares the base's components and
  releasing them would move them to the CPU underneath the pipeline being built from them.
• Keys are namespaced per library, so co-tenants cannot displace each other and ClearPipelineCacheNode
  empties only this library's.

The artifact still carries config_hash rather than becoming a cache reference: a reference is
resolve-or-raise, and the description is needed on the orchestrator too.

The builder's state override is removed. It reported the node UNRESOLVED when its pipeline was
missing from the cache, but node state is read on the orchestrator and the pipeline is held in the
worker, so it always answered "not cached" and re-ran the builder on every execution.
get_or_build_pipeline rebuilds on a miss in the process that holds the cache.

Verification

The diff against #66 is the object-store adoption only; #66 carries the execution-environment move and
the merges from main. make check and make test/unit pass.

[cjkindel]

@griptapeops review

[griptapeops]

✅ GriptapeOps finished @cjkindel's request in 8m 20s

• Read PR and diff
• Verify findings
• Post review and status

Review posted at 052247c.

[griptapeops]

The move to local_objects is aimed right, and the one-resident policy matches what ModelCache did. 1 correctness finding on the new Release Pipeline node, plus 1 LLMism.

• release_pipeline_node.py:49: releasing one pipeline breaks the other one in the base + ControlNet pair
• latent_diffusion_pipeline_builder_node.py:90: a comment describing code that was removed

[cjkindel]

@griptapeops re-review

[griptapeops]

✅ GriptapeOps finished @cjkindel's request in 1m 28s

• Read prior findings
• Check new commits
• Post re-review + status

Review posted at a34e105.

[griptapeops]

Both findings resolved. a34e105 removes the Release Pipeline node and the stale builder comment; the other commits since 052247c are merges from main via the base branch.