perf: Adjust coalesce for [<tiny range>, <massive range>]

[nameexhaustion]

In the existing logic we didn't coalesce this if the <massive range> was large enough that the distance to the chunk_size would be too big.

But I do think this is more of a nitpick as we probably don't encounter this case very often 😄

*EDIT: This causes a drop in performance compared to #19728, will be fixed in a later PR -

query	time
pl.read_parquet(path)	4.162399
pl.scan_parquet(path).select(cs.by_index(range(6_000))).collect()	2.437659
pl.scan_parquet(path).select(cs.by_index(range(0, 12_000, 5))).collect()	3.838982
pl.scan_parquet(path).select(cs.by_index(range(0, 12_000, 3))).collect()	3.886243

[nameexhaustion]

This previously gave -

[
    (0..1, 1), // <-- extremely small request
    (1..67108865, 0),
    (67108865..134217728, 2),
    (134_217_729..134_217_730, 3), // <-- extremely small request
    (134217730..201326593, 0),
    (201326593..268435456, 4),
]