perf(encode): add MarshalAppend for zero-alloc buffer reuse

[xe-nvdk]

Summary

• Adds MarshalAppend(dst, v) that appends encoded bytes to a caller-provided buffer, eliminating the final make+copy allocation in Marshal()
• Refactors Marshal() to delegate to MarshalAppend(nil, v) — behavior is identical
• Follows the standard Go Append* pattern (like strconv.AppendInt)

Benchmark Results

BenchmarkStructMarshal-14         3139424    378.4 ns/op   1224 B/op   4 allocs/op
BenchmarkStructMarshalAppend-14   4091446    293.8 ns/op     72 B/op   3 allocs/op

~26% faster, 94% less memory when callers reuse a buffer.

Test plan

• All existing tests pass
• New benchmark validates improvement