Optimize Capstone x86 disassembly with O(1) lookup tables

Replace binary search (log2(15K) ~14 comparisons) with direct lookup
tables for instruction mapping in the x86 Capstone decoder. Profiling
with callgrind showed find_insn() binary search was called 2-3x per
instruction and consumed ~22% of total disassembly cycles.

Changes in subprojects/capstone-v5:
- X86Mapping.c: Replace find_insn() binary search with O(1) direct
  index table (lazily built on first use, ~60KB memory)
- X86Mapping.c: Replace X86_insn_reg_intel/att binary searches with
  packed O(1) lookup tables
- SStream.c: Add fast integer-to-string formatters (fast_utoa_hex,
  fast_utoa_dec) replacing vsnprintf for number formatting
- SStream.c: Simplify SStream_concat0 overflow check
- MCInst.c: Use memset for tied_op_idx initialization instead of loop

Benchmark results (x86_64 disassembly, 65KB of /bin/ls, 50 iterations):
  cs_disasm:          2.03M -> 2.42M insns/sec (+19%)
  cs_disasm_iter:     2.43M -> 3.16M insns/sec (+30%)
  cs_disasm_iter (no detail): 3.63M -> 4.09M insns/sec (+13%)

Callgrind instruction count: 664M -> 525M (-21%)

Per-function improvements:
  X86_get_op_access:  95.1M -> 12.5M instructions (-87%)
  X86_get_insn_id:    50.8M -> 17.5M instructions (-65%)
  X86_insn_reg_intel: 30.1M ->  4.7M instructions (-84%)

All 1378 x86_64 asm tests pass (0 new failures).

https://claude.ai/code/session_01KDR9eBZ4vEAftFBQ2vuhmr

Author: claude

subprojects/capstone-v5.wrap

@@ -3,5 +3,5 @@ url = https://github.com/capstone-engine/capstone.git
 revision = 0bd647456e57ae17796937b9f296ef8bc1201b9e
 patch_directory = capstone-v5
 directory = capstone-v5
-diff_files = capstone-v5/capstone-patches/fix-x86-16.patch
+diff_files = capstone-v5/capstone-patches/fix-x86-16.patch, capstone-v5/capstone-patches/optimize-x86-lookup-tables.patch
 depth = 1