fix(model): refuse to remove a model/version still referenced by an endpoint (STO-360)

[jebenexer]

STO-360: https://linear.app/runpod/issue/STO-360/warn-about-dependent-endpoints-before-removing-model-repo-versions

• model remove now checks whether any serverless endpoint still references the model (or, with --hash/--version, that exact version) before removing it.
• If a dependent endpoint is found, the removal is blocked with a dependent_endpoints coded error naming the endpoint(s) — there is no override flag. A model/version in active use cannot be removed from the CLI.
• The fix path is spelled out in the error: detach it first with runpodctl serverless update <endpoint-id> --clear-models (or point it at a replacement with --model-reference <new-url>), then retry.
• The check queries GraphQL (myself { endpoints { modelReferences } }) rather than the REST endpoints list, since REST never returns modelReferences (verified live — it's simply absent from the wire shape).
• Matching is case-insensitive on the https://local/ scheme+host, since the backend persists it as https://LOCAL/... regardless of the case runpodctl emits it in (verified on both prod and RIAB).
• If the dependent-endpoint check itself fails (e.g. can't reach the API), the removal is blocked too — "can't verify it's safe" is treated the same as "not safe."

Also fixes a real, unrelated bug found while e2e testing this against a live endpoint: UpdateEndpointModels (backing serverless update --model-reference/--clear-models — the exact command this PR's error message tells users to run) always sent flashBootType: "" because REST GET /v1/endpoints/{id} returns a flashboot bool, never the flashBootType enum saveEndpoint expects, so every call failed enum validation with Value "" does not exist in "FlashBootType" enum.. Now derives the enum from the REST bool when the string is empty, same as serverless create already does.

Known, separate issue not fixed here: UpdateEndpointModels also drops GPU config — REST returns gpuTypeIds (array), but the struct only reads gpuIds (singular string), so saveEndpoint then fails with "gpuId(s) is required for a gpu endpoint" on any real GPU endpoint. This means serverless update --model-reference/--clear-models is still broken today for GPU endpoints created via serverless create --gpu-id. Worth its own follow-up ticket.

Test plan:

• go test ./...
• gofmt/go vet clean
• Live e2e verified against prod (api.runpod.io) and RIAB: blocked while a dependent endpoint exists (model/version untouched), proceeds cleanly once no endpoint references it, version-specific removal only blocks on the exact referenced hash (a different, unreferenced version removes cleanly), --force correctly rejected as unknown flag (no override exists).
• docs/runpodctl_model_remove.md regenerated for the new behavior/examples.

[promptless]

Promptless documentation updates

• Document runpodctl model commands, model-remove endpoint safety, and serverless update model flags documents this change in the runpodctl CLI reference. Docs PR