fix(index): narrow f16/f64 training chunks to f32 in streaming IVF trainers

[LuciferYang]

Problem

The three streaming IVF normalization sites called convert_to_floating_point, which returns f16 and f64 unchanged and only widens integer types. Those columns reached the trainers' unconditional Float32 downcast and panicked the build.

Fixes #8995.

What this changes

All three sites now go through one helper that casts the chunk to Float32. Two details worth calling out.

An f64 above f32::MAX saturates to an infinity rather than failing the cast. The samplers run their finite filter before the cast (ivf.rs:3193 and :3518), so without something else the trainers could receive infinities, and lance-index/src/vector/kmeans.rs:840 has no finite guard: the centroids would come out inf or NaN with no error. drop_rows_that_saturated removes exactly the rows the narrowing broke and logs how many, so a null row still passes through the way it does for every other value type.

The helper takes the array apart with into_parts rather than rebuilding the field by hand, which keeps the inner field's metadata and drops an unreachable branch. Its accept list is f16/f32/f64/i8, matching what a vector column can actually hold (rust/lance/src/index/vector/utils.rs:244).

The num_partitions <= 256 streaming path dispatches on value type already and is untouched.

Note for whoever reviews #8610: that PR moves this same convert_to_floating_point call into a new helper without changing it, so it carries the panic forward. Landing this first, or folding it in, avoids that.

Test plan

Five unit tests on the helper: the dtype matrix compared value by value, a sliced chunk mapping to the rows it points at, nulls staying on their row, an f64 that overflows f32 being dropped while a null row beside it survives, and an unsupported type returning InvalidInput with the type named. Removing the saturation filter makes the overflow test report [[1.0, 2.0], [inf, 4.0], [5.0, 6.0]].

Two end-to-end cases train a 257 partition model with streaming enabled, one f16 and one f64. They take different routes on purpose: a non-nullable column gets the fixed-range sampler, a nullable one the resampling path, and each has its own normalization site, so the two cases cover two of the three sites rather than differing only in dtype. Restoring the old normalization at the coreset site makes both panic in arrow's cast.

• cargo test -p lance --lib 3387 passed, 3 ignored
• cargo test -p lance --lib index::vector::ivf::tests:: 51 passed
• cargo clippy --all --tests --benches -- -D warnings clean
• cargo fmt --all --check clean

[lance-gatekeeper]

✅ Gate recommendation: approve.

This restores the streaming trainer’s intended Float32 boundary at the three f32-only coreset/refinement consumers while keeping native normalization and the ≤256 typed path unchanged. Post-cast overflow filtering preserves the finite-input invariant, and the regression coverage exercises both sampler routes and raw refinement.