Skip string coordinates in partition bound computation

[ghostiee-11]

Description

String-typed coordinates (e.g. station IDs, model names) crash during dataset registration because _block_metadata() calls np.min()/np.max() on the coordinate array, and numpy's minimum ufunc does not support string dtypes.

import numpy as np
import xarray as xr
from xarray_sql import XarrayContext

# Weather station observations with string station IDs
ds = xr.Dataset(
    {"temperature": (["station", "time"], np.random.rand(4, 10))},
    coords={
        "station": ["KJFK", "KLAX", "KORD", "KATL"],
        "time": np.arange(10),
    },
)
ctx = XarrayContext()
ctx.from_dataset("obs", ds.chunk({"station": 2, "time": 5}))
# ufunc 'minimum' did not contain a loop with signature matching types
# (dtype('<U4'), dtype('<U4')) -> None

Fix

Skip coordinates with string/object dtypes (kind in ('U', 'S', 'O')) before computing min/max bounds. The Rust ScalarBound enum only supports Int64/Float64/TimestampNanos, so string bounds would fail downstream regardless. Skipping makes the pruning logic treat the dimension conservatively (never prunes on it).

Also flattened the indentation by converting the if len > 0 block to an early continue.

Tests

Added test_string_coordinates. All 11 tests pass.

Partial fix for #121.

[alxmrs]

LGTM. I'm sorry, I introduced a merge conflict here by merging #155. But, that should be a quick fix, then I think this would be good to merge into main.

I really you filing the bug and making the fix! Welcome to the project.

[alxmrs]

Thanks, this is a good idea.

[alxmrs]

Thanks, good to have this test, and it's clear to read.

[ghostiee-11]

Thanks for the review!!

[ghostiee-11]

Fixed the conflicts..