When sampling deep call stacks repeatedly, most of the stack is typically unchanged between samples. This implements a per-thread frame cache that stores (address, frame_info) pairs. On each sample, if the current_frame has reached the last_profiled_frame pointer (indicating unchanged stack below), we stop walking and append cached frames instead of re-reading them from the remote process.

The cache is keyed by thread ID and stores the full frame list from the previous sample. When process_frame_chain hits the cached boundary, frame_cache_lookup_and_extend appends the cached continuation. After each sample we update last_profiled_frame in the target process so the eval loop can maintain the invariant.

This is opt-in via cache_frames=True on RemoteUnwinder to avoid changing behavior for existing users.