Make ring buffer size configurable at compile time via `LATTE_BUFFER_PWR`

[fior512]

Motivation

The buffer size is hardcoded as BUFFER_PWR = 16 (65 536 samples per ID per thread). This is a poor fit for two common situations:

• Embedded / constrained systems: 65 536 × 8 bytes = 512 KB per ID per thread is too much. A game running 10 IDs on 8 threads uses 40 MB just for Latte.
• Long-running stress tests: users want to capture millions of samples without wrap-around, requiring a larger buffer.

Proposed Change

// User defines before including the header (optional):
// #define LATTE_BUFFER_PWR 20   // 1M samples

#ifndef LATTE_BUFFER_PWR
  #define LATTE_BUFFER_PWR 16   // default: 65 536
#endif

static constexpr size_t BUFFER_PWR      = LATTE_BUFFER_PWR;
static constexpr size_t MAX_SAMPLES     = 1ULL << BUFFER_PWR;
static constexpr size_t BUFFER_MASK     = MAX_SAMPLES - 1;

A static_assert(LATTE_BUFFER_PWR >= 4 && LATTE_BUFFER_PWR <= 24, ...) guards against nonsensical values.

Acceptance Criteria

• LATTE_BUFFER_PWR respected when defined before #include "Latte.hpp"
• Default behaviour unchanged (power = 16)
• static_assert with a clear message for out-of-range values
• Memory usage table added to README (power → bytes per ID per thread)
• speedtest.cpp and testcase.cpp still compile and pass at default power