Fix backtraces from syscall stubs

[mkeeter]

Fixes oxidecomputer/hubris#2236

Functions defined with naked_asm do not have unwind info (see also rust-lang/rust#146736). For tasks which are in a syscall stub, this prevents our stack unwinder from making any progress: it can't get the unwind info for the bottom stack frame, so can't walk up the stack at all.

This PR adds a new function stack_syscall which assumes that we're at a syscall, and begins stack walking one frame up to escape the tarpit: we use LR as the program counter, and populate register values based on the pushes at syscall entry. We then prepend a synthetic stack frame with details about the syscall stub, since we can get its name from ELF tables.

stack_syscall is called if the normal hubris.stack(..) call fails, but we did get registers (indicating that there wasn't a stack overflow). In practice, it blends in well:

$ humility -d humility-bin/tests/cmd/cores/hubris.core.new-compiler tasks -sl user_leds
system time = 6611
ID TASK                       GEN PRI STATE
 2 user_leds                    0   5 recv, notif: timer
   |
   +--->  0x240019a0 0x080091dc userlib::sys_recv_stub <-- this is the synthetic stack frame
          0x24001a00 0x08009323 userlib::sys_recv
                     @ /hubris/sys/userlib/src/lib.rs:320:8
          0x24001a00 0x08009323 idol_runtime::dispatch
                     @ /git/idolatry-25cdcedc9b4829b2/f67fffe/runtime/src/lib.rs:237:20
          0x24001a00 0x08009323 main
                     @ /hubris/drv/user-leds/src/main.rs:206:9

[labbott]

bail vs anyhow?

[mkeeter]

I think bail is correct, because it builds + returns the error. anyhow just builds the error, so it would be more verbose (return anyhow!(..)).

[labbott]

Ahh yeah I think I have convinced myself that bail actually means "exit the program with this error" which is not actually true because we actually handle the error with the .or_else above