[Story]: Rework how cudf-polars resolves executor/engine configuration

[Matt711]

cudf-polars resolves executor/engine config (explicit value, then env var, then default)
in three separate places today: StreamingOptions, StreamingExecutor/
ConfigOptions.from_polars_engine, and a handful of ad hoc reads inside
SPMDEngine/DaskEngine/RayEngine.__init__. The same env var name and default end up
written two or three times for the same field, with nothing keeping them in sync. Some
fields end up reading a hardcoded fallback instead of the actual env var because one of
the duplicated copies never checks it, and _reset() doesn't consistently preserve a
field's previous value across an unrelated reset.

We want one shared way to resolve a field's value regardless of which class declares it,
only deferring resolution for fields that truly need it, and resolving an engine's config
once at construction instead of on every query.

Sub-issues:

• StreamingOptions._from_argparse silently drops env vars #23741 StreamingOptions._from_argparse silently drops env vars
• _reset() reverts unrelated executor_options fields to their default #23742 _reset() reverts unrelated fields to their default
• unify env var/default resolution into shared primitives #23743 Unify env var/default resolution into shared primitives
• narrow UNSPECIFIED to fields that actually need to defer #23744 Narrow UNSPECIFIED to fields that actually need to defer
• resolve engine-bootstrap options once, at construction, and freeze them #23745 Resolve engine-bootstrap options once, at construction, and freeze them
• dynamic_planning/join_filter_pushdown env var resolution lives outside _opt() #23754 dynamic_planning/join_filter_pushdown env var resolution lives outside _opt()
• Options doesn't respect a RAPIDSMPF_* env var unless the caller forwards it rapidsai/rapidsmpf#1159 Options doesn't respect a RAPIDSMPF_* env var unless the caller forwards it