fix(account_parameter): cast FLOAT-typed parameters to a Python float

[noel]

Summary

snowcap apply reported a phantom ~ UPDATE for INITIAL_REPLICATION_SIZE_LIMIT_IN_TB on every run, with the diff table showing identical 10.0 -> 10.0.

Root cause: SHOW PARAMETERS IN ACCOUNT reports this parameter's type column as FLOAT (confirmed live), not NUMBER. _cast_param_value only handled BOOLEAN/NUMBER/STRING, so it fell through to the else: return raw_value branch and the fetched value stayed a raw string ("10.0"). The YAML-declared value parses via PyYAML into a Python float (10.0). The diff in blueprint.py compares these with a plain !=, so "10.0" != 10.0 registers as a real change, while the printed table stringifies both sides independently for display, hiding the type mismatch.

_cast_param_value is the single shared function behind every SHOW-PARAMETERS-driven fetch (account, warehouse, table, task, and user parameters), so this fixes the same latent bug class for any FLOAT-typed parameter across all of those, not just this one.

Fix

Added a FLOAT branch to _cast_param_value that parses the raw string to a Python float, mirroring the existing NUMBER branch's error handling.

Test plan

• Added test_float / test_invalid_float_raises to TestCastParamValue, and an end-to-end TestFetchAccountParameter test reproducing the exact reported repro (SHOW PARAMETERS row typed FLOAT).
• Verified the new tests fail against the pre-fix code ('10.0' == 10.0 assertion error) and pass after.
• Full non-integration suite passes: python -m pytest tests/ --ignore=tests/integration (2218 passed, 1 skipped, 1 xfailed).
• make lint (black, codespell, ruff) and make typecheck (mypy) pass on tracked files.
• Ran ponytail-review (no findings) and code-review (no findings).

[github-actions]

Review of PR #76

No issues found. The change adds a FLOAT branch to _cast_param_value (snowcap/data_provider.py:652-656), mirroring the existing NUMBER/BOOLEAN/STRING branches in the same function, so SHOW PARAMETERS-reported FLOAT-typed values (e.g. INITIAL_REPLICATION_SIZE_LIMIT_IN_TB) are cast to Python float instead of falling through to the raw string in the else branch (data_provider.py:659-660), which previously caused spurious plan diffs against YAML-declared float values.

• Correctness: matches the pattern of the adjacent NUMBER branch and raises a clear exception on unparseable input, consistent with existing error handling.
• Tests: tests/test_data_provider.py adds direct unit coverage for _cast_param_value("10.0", "FLOAT") and the invalid-input path, plus an integration-style regression test through fetch_account_parameter that mocks execute and asserts the returned value is a float. Good coverage for the fix's actual failure mode (str vs. float causing drift in blueprint.py diffing).
• No SQL generation, privilege handling, or resource-architecture changes involved — this is purely a data-parsing fix confined to _cast_param_value.