Add cache_layer_indices to reduce wavefunction storage by selecting specific slice depths

[HaoranLMaoMao]

Problem

When cache_levels=["slices"] is set, MultisliceCalculator stores exit-wavefunctions for all nz propagation slices. For a typical simulation (440 slices × 4,096 MD frames × 25 probes × complex128) this produces ~4 TB per block. Users who only need EELS spectra at a few target thicknesses are forced to store and manage tens of terabytes of intermediate data, even though only a small fraction is ever used downstream.

---

Solution

This pull request adds a single optional parameter cache_layer_indices to setup():

ms.setup(
    cache_levels=["slices"],
    cache_layer_indices=[43, 87, 175, 263, 351, 439],  # None = all layers (default)
)

When set, only the listed layers are FFT'd and stored in frame_data / wavefunction_data; the remaining layers are discarded immediately after propagation without any FFT or disk write. The full multislice propagation through all nz slices still runs (physically required). WFData.layer is updated to hold the actual layer indices instead of arange(nz).

---

Changes

• MultisliceCalculator.setup(): new parameter cache_layer_indices: Optional[List[int]] = None
• MultisliceCalculator.run(): introduces self._active_layers to replace the hardcoded range(self.n_layers) loop; allocation of wavefunction_data and frame_data uses len(self._active_layers) as the last dimension
• WFData.layer: now holds the actual slice indices (e.g. [43, 87, 175, 263, 351, 439]) instead of arange(nz), so downstream code can map a target depth back to its compact storage slot via list(wave.layer).index(layer_idx). When cache_layer_indices=None, wave.layer remains identical to the original arange(nz).
• Fully backward-compatible: cache_layer_indices=None (default) preserves existing behaviour exactly — no changes required to existing scripts

---

Downstream usage example

Users specify layer indices directly. The corresponding physical depths can be
derived from slice_thickness for reference:

slice_thickness = 1.0   # Angstrom, same value passed to ms.setup()
target_layers   = [43, 87, 175, 263, 351, 439]
 
# Show what physical depth each layer corresponds to
for layer_idx in target_layers:
    depth_nm = (layer_idx + 1) * slice_thickness / 10
    print(f"  layer {layer_idx:4d}  ->  {depth_nm:.2f} nm")
# layer   43  ->  4.40 nm
# layer   87  ->  8.80 nm
# layer  175  -> 17.60 nm
# layer  263  -> 26.40 nm
# layer  351  -> 35.20 nm
# layer  439  -> 44.00 nm
 
ms.setup(
    cache_levels=["slices"],
    cache_layer_indices=target_layers,
    ...
)
wave = ms.run()
# wave.layer -> array([43, 87, 175, 263, 351, 439])
 
# Post-processing: map layer index to compact storage slot
for layer_idx in target_layers:
    out_idx = list(wave.layer).index(layer_idx)
    tacaw = TACAWData(wave, layer_index=out_idx)

---

Impact

For the amorphous-Si benchmark (nz = 440, saving 6 layers):

	Before	After
wavefunction_data shape	(25, 4096, 111, 111, 440)	(25, 4096, 111, 111, 6)
Disk per block	~4 TB	~55 GB
Saving	—	~98.6%