Add binary writer for sampling profiler

Implements streaming binary output with delta compression. The writer
tracks per-thread state to encode stack changes efficiently: identical
stacks use RLE, similar stacks encode only the differing frames.

String and frame deduplication uses Python's hashtable implementation
for O(1) lookup during interning. The 512KB write buffer amortizes
syscall overhead. When zstd is available, data streams through
compression before hitting disk.

Finalization writes the string/frame tables and footer, then seeks
back to update the header with final counts and offsets.

Author: pablogsal