perf: Decode() type switch for common scalar pointers

[xe-nvdk]

Problem

On the decode side, common patterns like `Unmarshal(b, &myString)` or `Unmarshal(b, &myInt64)` go through `reflect.ValueOf(v)` + full decoder dispatch. There is already a fast path for `*interface{}` but not for other common pointer types.

Proposal

Add fast paths in `Decode()` for:

• `*string`
• `*int64`, `*int`
• `*uint64`, `*uint`
• `*float64`
• `*bool`
• `*[]byte`

Each avoids `reflect.ValueOf` boxing + `getDecoder` sync.Map lookup.

Files

• `decode.go` — `Decode()` method

Expected Impact

MEDIUM — depends on how often Arc decodes into scalar pointers vs structs. Eliminates 1 reflect.ValueOf + 1 sync.Map load per scalar decode.