[tristan-deep]

Makes reconstruct.py run on a whole ulmshare acquisition (76000 frames). Before, it held every frame in memory (~80 GB of IQ) and was killed.

• Per-buffer processing: each 400-frame buffer is streamed, beamformed, SVD-filtered, localized and tracked, then dropped. Power Doppler, movie and ULM tracks are accumulated as it goes. The next buffer is read on a background thread. Peak memory stays flat at ~6 GB whatever N_FRAMES is. A full acquisition takes ~40 min.
• Stage checkpoints: each buffer's beamformed IQ, Power-Doppler stats, localizations and tracks are saved under zea's cache directory (CHECKPOINT_DIR). Each key covers the file, the buffer, every relevant setting and the source code of the functions that compute the stage. An interrupted run resumes, and a changed ULM parameter recomputes only the stages it affects. The beamformed stage takes ~1 MB per frame; CHECKPOINT = False turns checkpoints off.
• Axial upsampling: the IQ data is sampled at 5.2 MHz (148 µm per sample, ~6 grid pixels), and linear interpolation in delay-and-sum made the envelope dip between samples. That locked ULM localizations into rows and smeared the B-mode. The channel data is now upsampled 4× along depth by FFT zero-padding before beamforming (AXIAL_UPSAMPLING).
• ulm.py: localization is vectorised over the whole buffer, with results identical to the per-frame loop.

Output changes: the B-mode now shows the skull and brain speckle instead of streaks. Power Doppler loses its faint horizontal banding. ULM uses a per-buffer detection threshold and no longer links tracks across buffer boundaries, as the MATLAB reference does. The full-acquisition ULM map still shows an angled hatching pattern; that isn't fixed here.

Tested with zea 0.1.6.

Files changed: 2.

~ datasets/ulmshare/reconstruct.py
~ datasets/ulmshare/ulm.py