pef: Lazy metadata access

[srivarra]

Why: open_ome_zarr() on HCS plates had a throughput regression caused by two things in the metadata parsing path:

1. _first_pos_attr() used next(group.groups()) to find the first position's channel_names/axes, but zarr v3's groups() eagerly enumerates all children before yielding the first; scaling linearly with FOVs/well (0.2ms for 1 FOV, 259ms for 2500)
2. unique_validator() built pandas DataFrames just to check field uniqueness

Changes:

• Made axes and channel_names lazy & cached via functools.cached_property on NGFFNode — they resolve on first access, not during __init__
• On Plate, overrode both with cached_property that uses direct path access (self.zgroup[well_path][pos_name]) instead of groups() traversal
• Replaced unique_validator's model_dump() + pandas DataFrame with a simple set length check, removed the pandas import

Before & After: on some test data

Plate	Wells	FOVs/well	Before	After	Speedup
tiny	1	1	3.3 ms	0.63 ms	5.2x
few_wells	6	1	3.6 ms	0.71 ms	5.1x
96w_single	96	1	7.2 ms	0.78 ms	9.2x
96w_tiled	96	25	12.0 ms	0.90 ms	13.3x
25w_100fov	25	100	25.2 ms	0.68 ms	37.1x
384w_single	384	1	11.3 ms	1.13 ms	10.0x
dense_tile	1	2,500	535.8 ms	0.67 ms	799x

[gav-sturm]

This looks good to me. Has my okay and thanks for such a rapid turnaround. Will benefit a lot of our runs.