[r4victor]

• Check all of core/backends with pyright, excluding oci (whose SDK is costly to analyze), with two line-level ignores in base/offers.py until a gpuhunt release with float QueryFilter disk size bounds (Annotate QueryFilter disk size bounds as float gpuhunt#281).
• Raise the Node heap limit for pyright in CI to 4 GB, since the default ~2 GB heap on the macOS runners is no longer enough.
• Fix pyright errors in the cloudrift, lambdalabs, vultr, gcp, nebius, and azure backends and in base/offers.py, mostly by narrowing optional values whose presence is guaranteed.
• Make the Lambda API client raise on any non-OK /instance-types response instead of returning None.
• Fix Nebius credential errors passing field names instead of field paths to raise_invalid_credentials_error.
• Validate that stored Azure configs have resource_group and read it with AzureConfig.get_resource_group() instead of narrowing the field to str in AzureStoredConfig.
• Drop the unused tenant ID from the return value of Azure auth.authenticate().
• Ignore pyright errors that come from SDK typing: GCP compute_v1 enum members (inferred as ints in the huge generated module), the Nebius SDK Operation failing its own OperationMessage bound, and the flattened kwargs of Azure VM model constructors that are only missing from the TypeSpec-generated overloads.