Call `resetDiscoveredInputs()` on `unused_inputs_list` actions to free the flattened set of used inputs.

The flat `NestedSet` of used inputs is memory-inefficient because the original `NestedSet` structure is lost. Note that this change does not save memory on builds with `--notrack_incremental_state` or `--nokeep_state_after_build` because those already call `resetDiscoveredInputs()`.

In the new implementation:

* New method `Action#prunedInputs` communicates whether the action detected unused inputs.
* `resetDiscoveredInputs()` is called after execution to return the action to a pre-input-discovery state.

Other changes:

* The map of inputs in `EnhancedStarlarkAction#afterExecute` is initialized lazily to reduce garbage when there are no unused inputs.
* `ActionCacheAwareAction` is no longer necessary because `prunedInputs()` tells `ActionCacheChecker` to store all inputs' exec paths in the action cache.
* `ActionExecutionFunction` calls `prunedInputs()` to determine whether to request action inputs up front or wait until input discovery[*]. Actions that pruned inputs can get an action cache hit without requesting unused inputs.
* `addDiscoveredInputs()` uses `ActionInputMapHelper` to handle aggregation artifacts, which can now be "discovered" (when we avoid requesting the inputs up front). This is cleaner anyway.

The only behavior change is that `aquery --noinclude_pruned_inputs` still reports unused inputs, even after a build. This seems like an acceptable compromise for the memory savings.

[*]The initial execution of the action always requests the full set of inputs up front in order to take advantage of `ArtifactNestedSetKey` dependency structure. Discovered inputs are always requested directly, not via `ArtifactNestedSetKey` (although I don't know of a fundamental reason why).

PiperOrigin-RevId: 858177876
Change-Id: I72a952e70c7afe8341185304434f1424c827ca05