Reduces CPU consumption when profiling is enabled

Added the following functions:

• margo_push_prefinalize_callback
• margo_pop_prefinalize_callback
• margo_provider_push_prefinalize_callback
• margo_provider_pop_prefinalize_callback

• Adds RPC breadcrumbs
• Corrects a bug with provider-specific finalize callbacks

• Adds support for provider-specific finalize callbacks
• Adds margo_deregister

• Adds support for provider-based timed forward
• Adds support for asynchronous timed forward
• Replaces definition of margo_request with opaque pointer

Mainly corrects a bug where hg_handles were destroyed too early.