feature: implement all float, conversion, table & bulk memory opcodes with IEEE 754 hardening

Complete implementation of remaining WASM MVP and post-MVP opcodes:

Float arithmetic (f32/f64):
- abs, neg (purely bitwise), ceil, floor, trunc, nearest, sqrt
- add, sub, mul, div with indeterminate form handling
  (Inf+(-Inf), Inf-Inf, Inf*0, 0/0, Inf/Inf -> NaN; x/0 -> ±Inf)
- min, max with signed-zero fix (OR for min, AND for max)
- copysign (bitwise sign transfer)

Float comparisons (f64 eq/ne/lt/gt/le/ge):
- Added bitwise NaN detection before comparisons to prevent EInvalidOp on x64 FPC

Conversion opcodes:
- i32.wrap_i64, i64.extend_i32_s/u
- i32/i64.trunc_f32/f64_s/u with NaN + overflow traps
- f32/f64.convert_i32/i64_s/u
- f32.demote_f64, f64.promote_f32
- i32/i64.reinterpret_f32/f64, f32/f64.reinterpret_i32/i64
- i32.extend8_s, i32.extend16_s, i64.extend8_s/16_s/32_s

0xFC prefix dispatch + saturating truncation (FC 00-07):
- i32/i64.trunc_sat_f32/f64_s/u with NaN->0, clamp to min/max

Bulk memory (FC 08-0B):
- memory.init, data.drop, memory.copy (overlap-safe), memory.fill

Table operations (FC 0C-11 + opcodes 0x25/0x26):
- table.get, table.set, table.init, elem.drop
- table.copy (overlap-safe), table.grow, table.size, table.fill

Control flow & infrastructure:
- select t* (typed select, 0x1C) with LEB128 type vector
- block/loop/if blocktype now parsed as LEB128 s33 (multi-byte support)
- Control flow scanner updated for all new opcodes (0xFC sub-opcodes,
  table.get/set, typed select)
- VM init: jump tables heap-allocated via kalloc, FC table initialized

All float ops use bitwise NaN detection to avoid FPC x64 EInvalidOp.

.gitignore

@@ -1,3 +1,4 @@
 bin
 lib
 AGENTS.md
+GAP_ANALYSIS.md