fix(index): check quantizer params before sampling training data

[LuciferYang]

Closes #9367

load_or_build_quantizer returns early when a quantizer is already set, so everything below that runs with self.quantizer == None. The tail match &self.quantizer could therefore only take its None arm, and that arm held the only check on quantizer_params — a builder given neither params nor a quantizer sampled the training data and computed residuals before failing. The params are now bound once at the top with let ... else, the tail collapses to the Q::build call, and sample_size_hint reads from that binding.

The None => 256 * 256 sample-size default went with it. Its comment says "here it must be retrain", but retrain arrives with quantizer: Some(..) and returns at the top, so that arm was unreachable; the case it was written for ended in the params error either way.

No behaviour changes for any reachable input. The params-missing path is only reachable by calling IvfIndexBuilder::new(.., None, ..) explicitly — the one constructor that leaves quantizer_params empty, new_incremental, sets quantizer at the same time — and that path produced the same error before, just later. So this is dead-code removal and a check moved earlier, not a fix for something a user can hit.

How was this patch tested?

No new test: there is no reachable input whose behaviour differs, so there is nothing new to pin. cargo test -p lance --lib index::vector covers the paths that do build a quantizer, including the incremental one that returns at the early exit.

[LuciferYang]

Added test_load_or_build_quantizer_checks_params_before_sampling. The subtlety is that both orders end in the same error message, so the test points the builder at a column the dataset does not have: with the check first the call reports the missing params, with the check back below the sampling it reports column no_such_column does not exist in schema. I confirmed that by moving the guard back and watching the test fail with exactly that message.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The follow-up closes the earlier coverage gap: the new test distinguishes validation-before-sampling from the old order by making sampling fail on an invalid column. The upfront guard is a focused fix and preserves the preloaded-quantizer path.