… and float precision bugs

- listvalidator.ExactlyOneOf across the four *_model_config blocks
  (attached once; the framework tallies every sibling even when the
  attached block is null) catches multi-member configs (which previously
  silently dropped members and produced a perpetual diff) and empty
  model {} (which previously surfaced an internal
  "always an error in the provider" message) at plan time.
- stringvalidator.RegexMatches on harness_name enforces the service
  pattern ^[a-zA-Z][a-zA-Z0-9_]{0,39}$ offline; names starting with a
  digit or containing dashes previously validated cleanly and only
  failed at CreateHarness.
- max_tokens is Optional+Computed with UseStateForUnknown. The Update
  API retains the existing value on omission ("if not specified, the
  existing value is retained"), so state absorbs the retained server
  value instead of showing a perpetual "2000 -> null" diff after
  removing max_tokens from configuration.
- temperature and top_p on all four model configs are modeled as
  Float64 (schema + model struct). AutoFlex's flattenFloat32 routes
  float32 -> types.Float64 through decimal.NewFromFloat32 (comment:
  "Avoid loss of equivalence"), so imports round-trip cleanly
  (temperature = 0.7 stays 0.7, not 0.699999988079071) and no
  ImportStateVerifyIgnore workaround is needed.

Add acctests: harness_name regex ExpectError, model-union ExactlyOneOf
ExpectError (multi-member and empty), and a max_tokens set->unset
transition asserting PostApplyPostRefresh ExpectEmptyPlan (no
perpetual diff).

Dropping the duplicated memory commit (hashicorp#48655's 7151e9d) resurfaces the base
retrieval_config.relevance_score (Float32), so the float32validator import is
still needed. The model-config float-precision (Float64) work stays intact.