feat: support source filters in sensor references for flex-model and flex-context

[Copilot]

Sensor references in flex-model and flex-context fields can now carry source filters (source-types, exclude-source-types, sources, source-account) to narrow which beliefs are used when querying a sensor. This enables use cases like rescheduling a PV asset without separate sensors for forecasts vs. schedules — source filtering masks the unwanted beliefs on the same sensor.

Source filter support

• Added source_account field to SensorReference; all four filter fields are now forwarded to TimedBelief.search in get_series_from_quantity_or_sensor
• VariableQuantityField._serialize now round-trips source filter fields (previously dropped them)
• _resolve_soc_at_start_from_sensor propagates source filters when inferring SoC from a filtered sensor reference
• storage-efficiency in the storage scheduler now accepts SensorReference (filtered sensor) in addition to plain Sensor

Validation (source filters rejected where meaningless)

• aggregate-power in FlexContextSchema: rejects SensorReference — this field identifies a meter, not a belief query
• consumption and production in StorageFlexModelSchema / DBStorageFlexModelSchema: reject source filters — these fields designate output sensors to write the schedule to, not belief queries

// ✅ Valid: filter SoC beliefs by source account
{
  "state-of-charge": {
    "sensor": 42,
    "source-account": [7]
  }
}

// ❌ Invalid: consumption/production are write targets, not belief queries
{
  "consumption": {
    "sensor": 42,
    "source-types": ["scheduler"]
  }
}

A module-level constant _SENSOR_REFERENCE_SOURCE_FILTER_KEYS is shared across all four source-filter validators to keep the key list in one place.

[BelhsanHmida]

@copilot please take another pass on this PR. The core source-filter lookup path is useful, but a few PR-introduced behaviors still need to be handled:

1. Fix state-of-charge filtering when soc-at-start is inferred.
   ensure_soc_at_start() runs before StorageFlexModelSchema turns raw JSON into SensorReference, so _resolve_soc_at_start_from_state_of_charge() currently drops the new filter fields in the raw dict branch. Please preserve source-types, exclude-source-types, sources, and source-account there.

2. Preserve filters when serializing SensorReference.
   VariableQuantityField._serialize() currently dumps a filtered reference as only {"sensor": id}. Please include the filter fields so source-filtered references can round-trip.

3. Handle filtered storage-efficiency.
   With this PR, filtered storage-efficiency deserializes to SensorReference, but storage runtime still only treats plain Sensor as sensor-backed efficiency. Please make this path work and keep the resolution conversion correct.

4. Clarify/fix aggregate-power.
   aggregate-power is an output target, so source filters probably should not be accepted there. If the intended behavior is to allow it, please unwrap to the underlying sensor before using it as an output key. Otherwise, reject filtered references for this field.

Please add focused regression tests for these paths, especially the soc-at-start inference case with two beliefs on the same SoC sensor from different sources.

[Flix6x]

@BelhsanHmida thanks again for the review and revisions. I code reviewed everything once more, except for the tests. Maybe you can take a final look at those, to make sure they have proper coverage (I sometimes still spot copilot creating tests that cover trivial cases only).

[BelhsanHmida]

@BelhsanHmida thanks again for the review and revisions. I code reviewed everything once more, except for the tests. Maybe you can take a final look at those, to make sure they have proper coverage (I sometimes still spot copilot creating tests that cover trivial cases only).

Looks good to me now. I re-checked the test coverage too, and the important filter paths are covered with non-trivial cases.