STO-691: fix serverless update --model-reference/--clear-models on GPU endpoints

[jebenexer]

Fixes STO-691.

Problem

serverless update <id> --model-reference <ref> and --clear-models both
failed on every GPU endpoint:

{"error":"failed to update model references: graphql error: gpuId(s) is required for a gpu endpoint","code":"graphql_error"}

Root cause

UpdateEndpointModels round-trips the endpoint's current config through
saveEndpoint, reading the GPU pool id from endpoint.GpuIDs. But REST
GET /v1/endpoints/{id} never returns gpuIds — only gpuTypeIds, an array
of gpu type names ("NVIDIA A40"), a different identifier space from the
pool ids saveEndpoint wants ("ADA_24"). So GpuIDs was always empty, and
saveEndpoint rejects an empty gpuIds on any endpoint without instanceIds.

Fix

When the REST read yields an empty gpuIds and a non-empty gpuTypeIds,
UpdateEndpointModels resolves the pool id via ResolveServerlessGpuPoolID —
the same helper cmd/serverless/create.go uses for --gpu-id. An unknown gpu
type now errors instead of sending an empty gpuIds on to saveEndpoint.

Also adds the gpuIds vs gpuTypeIds REST/GraphQL asymmetry to AGENTS.md.

Testing

• internal/api/endpoints_test.go: gpuTypeIds resolves to the right pool id
  in the saveEndpoint input; an unresolvable type fails loudly and never
  reaches the write.

• e2e/cli_test.go: TestCLI_ServerlessUpdateModelReferenceOnGPUEndpoint
  creates a real GPU endpoint and exercises --model-reference and
  --clear-models — the gap called out in the STO-360 test plan.

• go vet ./..., go build ./..., go test ./..., govulncheck ./...

• go test -tags e2e ./e2e (needs a live API key against dev.runpod.io)

---

Note

Medium Risk
Touches serverless endpoint updates that full-replace config via saveEndpoint; scope is narrow (GPU id resolution on model-reference/clear-models) but mistakes could mis-assign GPU pools or break updates.

Overview
Fixes STO-691: serverless update --model-reference and --clear-models no longer fail on GPU endpoints with gpuId(s) is required for a gpu endpoint.

Those updates round-trip the endpoint through GraphQL saveEndpoint after a REST GET. REST only returns gpuTypeIds (display names), not gpuIds (pool ids like ADA_24), so gpuIds was always empty on write. UpdateEndpointModels now maps gpuTypeIds → pool id via ResolveServerlessGpuPoolID (same as create) when gpuIds is missing; unknown types error instead of sending an empty gpuIds. The Endpoint struct gains GpuTypeIDs for the REST read shape, and AGENTS.md documents the asymmetry.

Unit tests cover successful resolution and unresolvable GPU types; a new e2e test creates a GPU endpoint and exercises both update flags.

^{Reviewed by Cursor Bugbot for commit c8e04d4. Configure here.}

[jebenexer]

@cursor review

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit c8e04d4. Configure here.}