Seeweb backend: pydantic v2 alignment and review fixes

[peterschmidt85]

Review fixes for dstackai#4070, targeting this fork's master so the main PR picks them up on merge.

Changes

• Migrate the configurator and backend-data models to pydantic v2 idioms (validate_extra_ignore, model_validate, model_dump_json) — dstack dropped the v1 shims, so __response__/parse_obj/.dict()/.json() now fail at runtime.
• Add the unallocated_resources parameter to get_offers_by_requirements (the ComputeWithFilteredOffersCached signature gained it; without it every offer listing raises TypeError).
• Update tests to gpuhunt's CatalogItem (RawCatalogItem was removed; provider is now required).
• Compose the first-boot script from the base helpers (get_setup_cloud_instance_commands + get_shim_pre_start_commands + get_shim_env) instead of slicing get_shim_commands()[:-1], and write /etc/dstack-shim.env from the env dict rather than env | grep.
• Align server/action statuses with the ecsapi SDK enums — notably adds the canonical Fail server status, which the previous set missed (a failed server would poll to the provisioning timeout instead of failing fast).
• Docs: control-panel section is Compute (not Calculation); drop the provisioning-internals paragraph; revert the /CI_TESTING.md .gitignore entry.

Verification

• pytest src/tests/_internal/core/backends/seeweb src/tests/_internal/server/routers/test_backends.py — 67 passed, 36 skipped; ruff clean.
• Live e2e on real Seeweb servers (L40S and RTX PRO 6000, both terminated after): systemd-managed shim comes up after the first-boot reboot, shim healthcheck reachable in ~5–13 min, CUDA 13 containers verified via NVIDIA forward compatibility (cuInit in-container on both GPUs), and NoCapacityError fallthrough exercised against sold-out plans.

🤖 Generated with Claude Code