Fix three EXIF reading bugs in exif_read.py

[ptpt]

Summary

Fixes three bugs in mapillary_tools/exif_read.py, each with a regression test in tests/unit/test_exifread.py.

1. Epoch GPS date is not rejected (extract_gps_datetime)

The "no GPS fix" placeholder date 1970-01-01 was meant to be discarded, but the guard compared a datetime against a date:

if dt is None or dt == datetime.date(1970, 1, 1):  # datetime == date is always False

datetime.datetime(1970, 1, 1) == datetime.date(1970, 1, 1) is always False, so the guard never fired and a bogus 1970-… timestamp was returned. Fixed by comparing dt.date().

2. Non-numeric SubSec crashes the read (parse_datetimestr_with_subsec_and_offset)

SubSecTime comes from untrusted image metadata. A non-numeric value raised an uncaught ValueError from int(subsec[:6]), aborting the entire capture-time read (the EXIF datetime path isn't wrapped by callers).

This now mirrors exiftool's behavior: use only the leading run of digits and ignore any non-numeric remainder; if there are no leading digits, the sub-second is simply not applied. Verified against exiftool 13.55:

SubSecTimeOriginal     : 12ab
SubSecDateTimeOriginal : 2019:01:01 12:00:00.12   # our reader returns .120000 too

3. Fractional GPS minutes truncated (parse_time_ratios_as_timedelta)

Hours and minutes were floored to int, so a fractional minute in the GPSTimeStamp rationals was silently dropped (e.g. 30.5 min lost 30 s). Now kept as float so the remainder carries into the total.

Cosmetic

ExifRead.extract_height passed "width" as the XMP reason label; corrected to "height".

Tests

Adds TestExtractGpsDatetimeFromEXIF (valid / epoch-rejected / fractional-minute, built end-to-end with PIL) and test_parse_subsec_uses_leading_digits (fully non-numeric ignored; partially-numeric "12ab" → .120000).

uv run pytest tests/unit/test_exifread.py -q
78 passed

ruff, usort, and mypy clean.