Decouple string_to_bytes from get_system_memory_stats

[julman99]

Surfaced by a kluster.ai review on PR #17 (see inline comment).

The observation

string_to_bytes in src/eatmemory.c is conceptually a pure conversion: take a string like "10M" and produce a byte count. The % branch breaks that abstraction by reaching out to the OS for get_system_memory_stats() to resolve "X% of available memory" into an absolute byte count.

} else if (unit == '%') {
    struct system_memory_stats memory_stats;
    get_system_memory_stats(&memory_stats);
    if (!memory_stats.supported) {
        *error = EM_ERROR_PARSE_INVALID_UNIT;
        return 0;
    }
    numerator = memory_stats.free;
    denominator = 100;
}

Why this is a design smell, not a bug

The function does the right thing today. The objections are:

• Couples parser to runtime environment. A pure parser is easier to reason about, easier to unit test against fixed inputs, and doesn't need to know the OS exists.
• Mixes layers. Parsing user input ("transport-layer logic") and resolving a percentage against current free memory ("business logic") are different concerns currently fused.
• Hardens testability. With the current shape, the only way to test the % path is to actually have a working get_system_memory_stats for the host platform; mocking it requires intrusion.

Possible shapes

string_to_bytestakes asystem_memory_stats*parameter (caller provides; can be NULL on platforms where it's not supported, in which case%` is rejected).