[terrelln]

The x86-64 Huffman decoding loops used %rbp to hold bits[0], so while they ran %rbp didn't point to a valid frame. Profilers that unwind with frame pointers, such as perf record --call-graph=fp, couldn't unwind past these functions, so every sample inside them lost its callers.

Set up a standard frame (push %rbp; mov %rsp, %rbp) and keep %rbp intact for the whole function:

• bits[0] moves from %rbp to %r8.
• ip[0], which used to live in %r8, now lives on the stack. It is loaded into %r15 only around RELOAD_BITS(0). %r15 holds olimit and var[0] elsewhere, and neither is live during RELOAD_BITS(0).
• In the 4X2 loop, DECODE stages the 2-byte symbol through %r15 instead of %r8, so %r8 no longer needs to be spilled around the decode step.

The CFI is updated to match. The CFA is defined relative to %rbp once the frame is set up, so the per-push .cfi_def_cfa_offset directives are dropped. On exit, the CFA switches back to %rsp after %rbp is popped.

Tested on x86-64 with BMI2, using gcc and clang, with and without -fcf-protection:

• tests/fuzzer and tests/zstreamtest pass.
• The CLI round-trips a range of inputs, levels and block sizes.
• gdb backtraces reach main from the prologue, loop body, and epilogue of both loops.
• With perf record --call-graph=fp, every sample inside the asm loops now unwinds to main. Before this change, none unwound past the leaf.
• Decompression speed is unchanged, benchmarked with zstd -b1e1r silesia/.