Summary

Customers can now ship their own QuerySolver (say, one that reshapes their raw
tables into the Impulse silver schema) without touching Impulse core.

Solvers register themselves by name, the report config picks one by that name through the
existing query_engine.solver field, and each solver can bring its own
SolverConfig subclass whose extra fields get checked at parse time.

Config picks a solver by registered name, not by class path. So a config can't
make Impulse import and run some arbitrary class. What solvers exist depends
entirely on what the driver imports.

DefaultSolver moves onto this mechanism with no behavior change. It registers
under "DefaultSolver" (plus the old "DeltaSolver" / "KeyValueStoreSolver"
aliases), so existing configs keep working.

Changes

• Registry (solvers/registry.py): @register_solver(name, config_cls, *, aliases, overwrite) plus resolve_registration / is_registered / registered_names. Maps a name to (solver_cls, config_cls), rejects conflicting duplicates, and lists the known names when a lookup misses.
• Uniform construction (QuerySolver.from_config + solvers/solver_context.py): a frozen SolverBuildContext (spark, solver_config, is_raw_data, drop_implausible_data, raw_encoder) is the one place construction inputs live. Base from_config passes just solver_config; DefaultSolver overrides it to also wire spark and the raw-data flags. The report factory builds any solver the same way, whatever its constructor looks like.
• Source-loading seams on QuerySolver: load_container_tags, load_container_metrics, load_channel_tags, load_channel_metrics, load_channel_mapping, load_unit_conversion, load_channels. Defaults read the single configured table (same as before); a custom solver overrides them to combine or reshape raw sources. DefaultSolver's reads now go through these seams.
• Config (config_parser.py): query_engine.solver is a plain str now (any registered name). Solvers becomes a StrEnum so old == Solvers.DEFAULT_SOLVER comparisons still hold. A mode="before" validator re-validates solver_config through the selected solver's config_cls, so a custom solver's fields are enforced at parse time and an unregistered name fails early as a ValidationError.
• Report factory (core/report.py): create_solver resolves the name through the registry and calls from_config(ctx) instead of a hardcoded match. Incremental change detection reads containers via solver.load_container_metrics(...), so a reshaping solver feeds the same container set into upsert detection.
• Docs (docs/impulse/docs/data_model/ingestion.md): a "Registering a custom solver" section covering subclassing DefaultSolver, declaring a SolverConfig subclass, selecting it by name, and importing the package so registration runs.

Test Plan

• Unit tests added/updated
• Manual testing completed
• Documentation updated (if applicable)

Checklist

• Code follows project style guidelines
• Self-review completed
• No new linter warnings introduced