Miscellaneous c18n improvements

[dpgao]

The PR bundles together three minor improvements to c18n.

1. When c18n is disabled, the _rtld_setjmp helper (and its counterparts for libunwind) must not access the trusted stack but should tail-call an epilogue assembly function.
2. Trampoline tables are now allocated with malloc instead of mmap.
3. Block all signals when a new thread is created and only restore the signal mask when the actual thread function is about to be called. This is need for full interrupted safety which will be implemented in the future.

[bsdjhb]

I think a smaller diff that is a bit easier to understand might be to add this at the top-level of the file:

/*
 * c18n: Always block all signals when creating a new thread to allow
 * RTLD to setup the environment to handle signals.
 */
#if defined(__CHERI_PURE_CAPABILITY) && defined(RTLD_SANDBOX)
#define CREATE_SUSPENDED(thr) 1
#else
#define CREATE_SUSPENDED(thr) ((thr)->attr.suspend == THR_CREATE_SUSPENDED)
#endif

And then replace the place that check for THR_CREATE_SUSPENDED with this new helper macro.

[dpgao]

Wouldn't that cause this line to run even though THR_CREATE_SUSPENDED is not actually set?

In my latest patch I just set create_suspended always to 1 when RTLD_SANDBOX is defined.