docs: clarify M4 support, long-video scaling, and GUI status

[Berserk-hub150]

Summary

Adds practical documentation addressing the parts of issue #5 that can be determined from the current implementation.

Changes

• clarifies that the current TensorOps/NAX path is M5-specific and that M4 uses the ordinary Metal/MPSGraph path
• explains that published M3/M5 timings should not be treated as M4 benchmarks
• documents memory considerations for large and long runs, especially the extra residency caused by --show
• explains why 15-second 768p runtime should be benchmarked at the exact target shape rather than extrapolated only from frames x pixels
• documents the current frontend status without making unsupported claims about future GUI plans

Validation

• git diff --check
• documentation-only change
• no invented M4 benchmark numbers

Related to #5.