Wrap gas counter (gas_left) in a strong Gas type

[chfast]

gas_left is a bare int64_t threaded through the whole interpreter, so it can be silently confused with unrelated integers (stack values, memory sizes, EVMC's own int64 gas fields). Wrap it in evmone::Gas: conversions to/from int64_t are explicit, so gas only crosses into raw int64_t at the EVMC boundary (message.gas, evmc::Result.gas_left, the GAS opcode push, tracing) through deliberate static_cast. The gas-domain operators (-=, +=, binary -, +, /, <=>) accept raw int64_t cost operands, so instruction bodies are unchanged and the change stays mechanical: Result::gas_left, check_requirements, every core impl, grow_memory/check_memory, dispatch[_cgoto] and
AdvancedExecutionState::gas_left.

Generated code is not identical to the int64_t version (measured on the clang assertions build, x86-64):

• baseline_execution.cpp.o gains 127 static instructions. The gas check (gas_left -= cost) < 0 lowers to subq; jl (signed-less, reusing the subtraction's flags) instead of subq; js (sign bit) -- same instruction count; in cmov contexts the spaceship form even drops a testq (subq; cmovge vs testq; cmovns).

• Hot path is zero-cost. Deterministic instruction counts (google benchmark --benchmark_perf_counters=INSTRUCTIONS) over single-opcode loops running millions of iterations stay within +0.003%: ADD +0.003%, MUL +0.001%, SUB/LT/GT/ISZERO/NOT/SIGNEXTEND/JUMPDEST +0.002..0.007%.

• A fixed ~255 extra instructions per execute() call remain (setup and teardown inlining churn from the boundary casts and the Gas dispatch return value): +0.003% on normal workloads, up to +0.9% only on trivially short executions (synth/loop_v1: 27544 vs 27289 instructions).

1115/1115 unit tests pass (baseline, advanced, cgoto).