Exception unwinding performance improvements

[dnadlinger]

To keep track of some things we could do to increase the performance of exception unwinding:

• Keep a free list (or just a constant-sized buffer) for _d_exception and cleanup block structs to avoid memory allocations.
• For the unwind target frame (i.e. where the catch is), cache the results of the search phase inside _d_exception (or the cache in the context on ARM EHABI). This spares us from iterating through the tables and comparing classinfos a second time.
• Mark up (re)throw-related runtime functions as nounwind. Might lead to slightly more compact call site code.
• Avoid pushing a cleanup record/calling _d_eh_begin_catch for landing pads that are only for catches and have no cleanup part. Might be a bit tricky to figure out, and requires putting the _d_eh_begin_catch call into a separate BB before calling the catch (which is shared between different landing pads), or otherwise handling this situation in a more intelligent way.