Preserve full DSP cycle counts

[RhythrosaLabs]

Summary

Resolves #31 by preserving the complete Blackfin cycle count end to end.

The change:

• widens t_profile.cycles to uint64_t
• removes the local truncating cast in knl_profile_stats()
• extends the DSP profile payload from 4 to 8 cycle bytes
• decodes all 8 cycle bytes on the CPU side
• rejects profile packets shorter than the new 12-byte payload

Validation

• Strict host checks pass for the standalone protocol round trip with 0x0123456789abcdef.
• git diff --check passes.
• Target build is blocked because the ARM/Blackfin cross-toolchains are not installed in this environment.

The complete implementation is also published on RhythrosaLabs/freetribe-studio main.