[cpu] support TF single-stepping to make gdb work

[lqs]

The trap flag (FLAG_TRAP) was previously logged as unsupported and cleared on update_eflags, so any guest debugger relying on single-stepping didn't work.

This adds proper TF handling: when TF is set, cycle_internal interprets one instruction at a time and raises #DB afterwards. DR6.BS is set so the guest can distinguish a single-step trap from a hardware breakpoint.

With this, gdb breakpoints and single-stepping (stepi/continue over breakpoints) now work inside the guest.

Fixes #278.

[copy]

Hi @lqs, thanks for the contribution. Just as a heads-up, it will be very difficult to get this merged. I will need the following:

• an assurance that all code here is entirely written by yourself (as opposed to code-generating llms). I don't review llm-generated code on principle
• handle trap-inhibit instructions (e.g. mov SS)
• handle jit-compiled code. It will need to bail out when popf or another instruction sets TF
• verify interaction with various system instructions (e.g. INT, IRET, HLT, STI, ...)
• test cases
• a benchmark that shows this doesn't negatively affect performance

[lqs]

Thanks for the review. Apologies, I missed this policy. The code was llm-generated, so I'll close the PR. Sorry for the noise.