fixed_params should accept keys required for runtime-supplied grids

[hmgaudecker]

Problem

broadcast_to_template (and the validator chain that follows) rejects
any fixed_params key that no DAG function lists in its signature.
That assumption breaks for runtime-supplied grids: a parameter
referenced by an IrregSpacedGrid(pass_points_at_runtime=True) (or
similar grid-spec consumer) is needed at grid-build time, never by a
DAG function — yet it still needs to flow through fixed_params so
the per-iteration value can size the grid.

The user-visible friction: callers either funnel the param through
model.simulate(params=...) (where it competes with estimation
parameters and arguably doesn't belong) or fabricate a no-op DAG
function that "consumes" it just to satisfy the validator.

Useful when some components of a grid change across optimizer
iterations but others are constant.

Proposal

Relax the validator so a fixed_params key referenced by any
runtime-supplied grid spec is accepted without a DAG-input footprint.
The grid spec already knows which params it needs; surface that to
broadcast_to_template.

[hmgaudecker]

Closing — the proposed solution (#351) is solvable on the user side instead.

Auxiliary scalars that only the grid construction needs (and that no DAG function consumes) don't have to round-trip through pylcm's params template. The user-side injector can take them as explicit Python kwargs:

def inject_consumption_dollars_points(*, params, model, max_consumption_dollars: float):
    pts = _compute_points(..., max_consumption_dollars)
    out = dict(params)
    for regime_name in model.regimes:
        out[regime_name] = {**out.get(regime_name, {}), "consumption_dollars": {"points": pts}}
    return out

This keeps broadcast_to_template's unknown-keys check intact (good for catching real param-typos) and avoids a new per-grid concept (extra_param_names). Migrating aca-model accordingly.