perf: overlap patch reads in split correction, honor percentiles

[anna-grim]

Split-correction feature extraction read image patches synchronously with the caller's forward pass, and computed all skeleton features before issuing the first read, so neither the GPU nor the network was kept busy.

• FeaturePipeline: launch the patch reads before computing skeleton features so the two overlap. Image features are stored after the skeleton pass, since proposal profiles are concatenated onto the skeleton-based proposal features.
• FragmentsDataset: use the prefetch parameter, which was previously unused, to extract features for upcoming batches on a background thread. This overlaps extraction with the caller's forward pass and bounds the number of resident batches. Lower the default to 2, since each queued batch now holds its image patches in memory.
• ImageFeatureExtractor: pin max_workers to 32, where read throughput plateaus, rather than inheriting a cpu-count-derived default.
• Pass percentiles through to the ImageConfig that ImageFeatureExtractor builds, which dropped the caller's value and normalized with the ImageConfig default. SplitInferenceConfig now carries the value, mirroring the existing patch_shape override.
• Store feature matrices as float32 rather than float64, halving the host-to-device transfer of image patches.

Note: split inference now normalizes with percentiles (1, 99.9), where it previously always used (1, 99.5) regardless of configuration.