fix(rust,python): make python schema_overrides information available to the rust-side inference code when initialising from records/dicts

[alexander-beedie]

Closes #12032.

• We weren't passing information available in the user-provided "schema_overrides" param down into the Rust-side functions where the initial type inference occurs; consequently large UInt64 values would first get inferred as Float64 before being cast back to UInt64, thereby losing accuracy.

• This PR enhances the Rust-side read_dicts and finish_from_rows methods, allowing them to integrate the additional/optional information so that they can establish the correct dtypes for overridden columns without first mediating via an inferred type.

Before

import polars as pl
recs = [
    {"id":9187643043065364490, "misc":333},
    {"id":9223671840084328467, "misc":666},
    {"id":9187643043065364505, "misc":999},
]
df = pl.from_records( recs, schema_overrides={"id":pl.UInt64} )

df
# shape: (3, 2)
# ┌─────────────────────┬──────┐
# │ id                  ┆ misc │
# │ ---                 ┆ ---  │
# │ u64                 ┆ i64  │
# ╞═════════════════════╪══════╡
# │ 9187643043065364480 ┆ 333  │   << all "id" values are wrong as
# │ 9223671840084328448 ┆ 666  │      they were cast back from f64
# │ 9187643043065364480 ┆ 999  │
# └─────────────────────┴──────┘

After

df
# shape: (3, 2)
# ┌─────────────────────┬──────┐
# │ id                  ┆ misc │
# │ ---                 ┆ ---  │
# │ u64                 ┆ i64  │
# ╞═════════════════════╪══════╡
# │ 9187643043065364490 ┆ 333  │   << all "id" values are now correct
# │ 9223671840084328467 ┆ 666  │      (avoided intermediate type/cast)
# │ 9187643043065364505 ┆ 999  │
# └─────────────────────┴──────┘

Note that if "schema_overrides" is not provided, the column is loaded as Float64, so the user should see that they need to provide an explicit override, where applicable.

---

🚀 As well as fixing this edge case, we should also get a little extra performance when loading from dicts where only partial type information is provided (via "schema_overrides"), as the affected columns will no longer trigger a post-init cast - they will be directly loaded as the correct type.

👀 Note: if the whole schema is given up-front (via the "schema" param) then none of these issues apply and everything is loaded correctly/efficiently.