Use approximate float comparison for JSON test records

[wavebyrd]

Fixes #196.

Python's shortest-repr float formatting varies across versions (and platforms), which causes the test recorder to detect spurious changes in JSON fixture files. For example, a value recorded as 0.011710000000000002 may serialize as 0.01171 in a newer environment -- both represent the same IEEE 754 double, just with different string representations.

What changed:

In tests/conftest.py, the Record class now falls back to math.isclose (relative tolerance 1e-9) when comparing JSON record files whose string representations differ. The exact-match fast path is tried first, so there is zero overhead when strings already match. Non-JSON records (CSV, txt) are unaffected.

How it works:

• _values_approx_equal recursively walks two JSON-parsed structures, using math.isclose for float-to-float pairs and strict equality for everything else (ints, strings, bools, nulls, structure/length).
• _json_strings_approx_equal wraps the above with json.loads and catches decode errors (falling back to exact comparison).
• record_changed checks the file extension and only uses the approximate path for .json records.

No recorded fixtures were changed. No new dependencies.