feat: support deferred index creation with WITH (train=false)

[ivscheianu]

Problem

ALTER TABLE ... CREATE INDEX always builds the index eagerly over all existing data before returning. On large datasets this is a long-running, blocking operation. There was no way to register an index definition without immediately training it, making it impossible to separate the "declare the index" step from the "populate it" step.

Solution

Add a train option to the WITH clause:

ALTER TABLE my_table CREATE INDEX idx_text_fts USING fts (text)
WITH (train=false)

When train=false, no Spark tasks are launched and no data is processed. An empty index is committed directly on the driver with an empty fragment bitmap — all existing rows appear as unindexed. A subsequent OPTIMIZE call covers them incrementally at a time of the caller's choosing.

The default (train=true) is unchanged, so all existing behaviour is preserved.

As a side fix, Spark-level execution options (train, build_mode, rows_per_range) are now filtered before being forwarded to the Lance index backend as index parameters.