feat: implement the URL search params serialization standard

[razor-x]

A Python port of @seamapi/url-search-params-serializer, exported for callers using their own HTTP client. Nothing in the SDK calls it yet — serializing the params of a request is a separate change.

The layer

The serialization defines the name and value of each search param, where every value is a string, and leaves rendering the query string to URLSearchParams. UrlSearchParams is that layer here, so serialize_url_search_params is a thin wrapper over it, exactly as the reference implementation wraps new URLSearchParams().

search_params = UrlSearchParams()
update_url_search_params(search_params, {"device_ids": ["device1", "device2"]})

list(search_params)   # => [('device_ids', 'device1'), ('device_ids', 'device2')]
str(search_params)    # => 'device_ids=device1&device_ids=device2'

Three places Python differs from JavaScript

These are where a naive port would produce the wrong bytes:

• Encoding. The application/x-www-form-urlencoded serializer leaves * bare and escapes ~. urllib's quote_plus does the opposite, so the encoder is written against the byte set.
• Sorting. URLSearchParams.sort() compares UTF-16 code units, so 😀 sorts before �. Python's str comparison sorts by code point and gets that backwards; the sort key is name.encode("utf-16-be").
• Numbers. repr(1.0) is "1.0" but JavaScript gives "1"; repr(1e-7) is "1e-07" vs "1e-7"; the exponent-notation thresholds differ too (Python switches at 1e16, JavaScript at 1e21). This implements the ECMAScript Number::toString algorithm on top of Python's shortest-round-trip repr.

Omitted params and null params

Python has one absence value and the standard needs both. Since sending null is rarely intended and unsetting a value cannot be undone, None keeps the safe meaning:

serialize_url_search_params({"name": NULL, "limit": 20})  # => 'limit=20&name='
serialize_url_search_params({"name": None, "limit": 20})  # => 'limit=20'

NULL is the value to pass; Null is its type, exported so a caller can annotate a wrapper. is_null is internal — the package does not export it.

Verification

Checked against the reference implementation running under Node:

• 116 hand-built cases covering every rule in its README, including error messages
• 4,000 randomized nested-param cases with unicode keys, mixed tuples, deep nesting, dates and bigints
• 60,000 float cases including every power of ten from 1e-330 to 1e308, denormals, and random binary64 bit patterns

All match byte-for-byte. Separately, 22 cases round-trip through the parser the API uses in strict mode, which is documented as a true inverse of the serializer.

The 54 unit tests in this PR are the readable subset of that: the ported test suite plus the Python-specific cases (naive datetimes, Decimal-free float formatting, dict subclasses, sets rejected as non-deterministic).

Checks

Rebased onto beta after #612, so this runs against the updated tooling: black 26, pylint 4, pytest 9 and mypy 2. 141 tests pass (87 already on beta, 54 added here), pylint 10.00/10, black, rstcheck, and mypy seam test clean across 114 files.

This PR is purely additive — six files, no changes to pyproject.toml, uv.lock, seam/client.py, codegen, or the generated routes.