fix(registry): merge duplicate model deployments instead of shadowing (#3709)

[nac7]

What

PR #3694 changed the deployment resolution order so that prod_env/model_deployments.yaml
entries are registered after the built-in helm.config entries, and
get_default_model_deployment_for_model picks [-1] (last registered) as
the default.

As a result, a prod_env entry that is incomplete — e.g. a user who
only wanted to change client_spec but left tokenizer_name unset —
silently shadows the full built-in entry. The WindowService then
receives tokenizer_name=None and raises:

ValueError: tokenizer_name is None

The confusing part: the user's config file is correct and a standalone
diagnostic script proves it loads fine, but helm-run still fails because
the prod_env entry wins.

Fix

src/helm/benchmark/model_deployment_registry.py — register_model_deployment

When a deployment name is already present in DEPLOYMENT_NAME_TO_MODEL_DEPLOYMENT,
merge the new entry over the existing one rather than blindly appending:

• The new entry's non-None fields override the existing values (preserving the intended prod_env-wins behaviour from Change logic for default model deployments #3694).
• Fields that are None in the new entry fall back to the existing value, so a sparse prod_env entry cannot accidentally wipe tokenizer_name (or max_sequence_length, etc.) that was set in the built-in config.
• name, client_spec, and deprecated are always taken from the new entry.
• A hwarn is emitted on every merge so that users can see which sources are competing.

# Before: blindly appended — prod_env None fields silently win
DEPLOYMENT_NAME_TO_MODEL_DEPLOYMENT[model_deployment.name] = model_deployment
ALL_MODEL_DEPLOYMENTS.append(model_deployment)

# After: merge non-None fields; None fields fall back to the existing entry
if existing := DEPLOYMENT_NAME_TO_MODEL_DEPLOYMENT.get(model_deployment.name):
    override_fields = {f.name: v for f in dataclasses.fields(model_deployment)
                       if (v := getattr(model_deployment, f.name)) is not None}
    override_fields |= {"name": ..., "client_spec": ..., "deprecated": ...}
    model_deployment = dataclasses.replace(existing, **override_fields)

Tests

src/helm/benchmark/test_model_deployment_registry.py (new file, 6 tests):

Test	Checks
test_first_registration_stored_as_is	Baseline: first entry stored unchanged
test_override_with_full_entry_replaces_all_fields	Full prod_env entry overrides everything
test_sparse_override_inherits_none_fields_from_builtin	Regression #3709: tokenizer_name=None in prod_env falls back to built-in value
test_sparse_override_none_max_sequence_inherits_from_builtin	max_sequence_length=None in prod_env falls back to built-in
test_deprecated_always_taken_from_new_entry	deprecated bool always uses new entry
test_only_one_entry_in_lookup_dict_after_override	Dict holds exactly one entry per name

Fixes #3709