Validate the hash operator's salt type instead of crashing on len()

[feiiiiii5]

What this fixes

The hash operator validates hash_type but never validates salt, so a salt of the wrong type survives validation and crashes inside operate():

TypeError: object of type 'int' has no len()

from presidio_anonymizer/operators/hash.py:41 (len(params[self.SALT])). Printed behaviour at f251c513, calling Hash.validate({"hash_type": "sha256", "salt": v}) directly:

  salt=123     -> TypeError (raw)
  salt=1.5     -> TypeError (raw)
  salt=True    -> TypeError (raw)
  salt=['x']   -> TypeError (raw)

This is not only an ugly exception type. Every sibling operator (replace, mask, encrypt) reports a bad parameter as InvalidParamError, and the anonymizer server turns that into a 4xx; a raw TypeError turns it into a 500 with no message for the caller. It also defeats the pre-flight check: validate() is the function the batch/EngineUtil path calls to reject bad configurations before any text is touched, and for hash it silently passed anything as long as hash_type was fine.

Change

The salt rules move into one _validate_salt() helper that both operate() and validate() call, so the two can no longer disagree:

• non-str/bytes/bytearray salt → InvalidParamError("Invalid salt type '<type>'. ...")
• str → encoded to bytes (unchanged)
• empty → the existing "Salt parameter cannot be empty..." message, verbatim
• shorter than 16 bytes → the existing "Salt must be at least 16 bytes (128 bits)..." message, verbatim
• None / absent → random per-entity salt, as documented

Behaviour changes this PR declares (per AGENTS.md "Declare behavior changes")

1. Hash.validate() now rejects a wrong-typed salt where it previously accepted it and let operate() blow up later. No previously-valid input becomes invalid.
2. An explicit "salt": null used to raise TypeError: object of type 'NoneType' has no len(); it now means "no salt supplied", i.e. the documented random per-entity salt. null is how JSON expresses an absent value, so this matches the operator's own docstring instead of contradicting it.
3. Error text for a wrong-typed salt is new; the two existing salt messages are untouched so anything matching on them still works.
4. 16 is now the constant MIN_SALT_LENGTH instead of a literal in two places; same value, same semantics.

Tests

Added to presidio-anonymizer/tests/operators/test_hash.py: wrong-type salts (int, float, bool, list, object) through validate() and through anonymize(... OperatorConfig("hash", {"salt": ...})), salt: null producing a random salt, and the two existing length messages still raised for str and bytes inputs.

Base control — operators/hash.py restored from f251c513e8aad6820e9b2a44983ebf3a93637b07, with git diff -- <that file> printing nothing so the run measures upstream code with the new tests kept:

$ python -m pytest presidio-anonymizer/tests/operators/test_hash.py -q
7 failed, 22 passed
FAILED ...test_when_salt_is_not_a_string_or_bytes_then_validate_raises[1234567890123456 / 1.5 / True / bad_salt3 / bad_salt4]
FAILED ...test_when_salt_is_missing_or_none_then_random_salt_is_used
FAILED ...test_when_salt_type_is_invalid_then_anonymize_raises_invalid_param_error

With the patch applied: 29 passed. Whole package: presidio-anonymizer/tests → 323 passed (excluding test_ahds_surrogate.py, which needs the optional presidio-ahds extra that is not installed here; that file is untouched by this diff).

Lint: ruff check on the two changed files reports the same 29 pre-existing findings at base as at head — this diff adds none. Every changed line is executed by the tests above, which is what the AGENTS.md ">= 90% coverage on changed lines" rule asks for.