Move dr_simd_t to separate allocation to support dynamic vector sizing and reduce stack usage

Xref discussion at https://github.com/DynamoRIO/dynamorio/issues/3544#issuecomment-2605265904 on splitting out dr_simd_t from the mcontext to help reduce offset sizes for RISC-V.

The primary cross-arch benefit of separating out dr_simd_t and making it dynamically allocated would be to support any vector size. Today both RISC-V and AArch64 support large SIMD vectors, but DR has a hardcoded cap (256-bit for RISC-V and 512-bit for AArch64) in order to keep the mcontext small and stack-allocatable.

The size of the mcontext is a problem on all platforms since DR uses small thread stacks. Allocating several mcontexts across a call chain can overflow the stack and with today's large SIMD fields inside mcontext we have to be careful declaring them on the stack. See comments in signal.c like these:
```C
/* It's safe to allocate since we do not send signals that interrupt DR.
* With priv_mcontext_t x2 that's a little big for stack alloc.
*/
si.mcontext = heap_alloc(dcontext, sizeof(*si.mcontext) HEAPACCT(ACCT_OTHER));
si.raw_mcontext = heap_alloc(dcontext, sizeof(*si.raw_mcontext) HEAPACCT(ACCT_OTHER));
```
```
/* Helper that takes over the current thread signaled via SUSPEND_SIGNAL. Kept
* separate mostly to keep the priv_mcontext_t allocation out of
* main_signal_handler_C.
```

However, there are signal handler paths where we need to declare and use new mcontexts and allocating heap is problematic. An indirected dr_simd_t could still be placed on the stack right after the mcontext: but only if there's stack space. We would have to create a new signal-safe heap pool for dr_simd_t, which has to be provisioned ahead of time and so has per-thread overhead to be considered (we run applications with thousands of threads).