We noticed that generating the sample format at the end of a profile can get
rather slow and this aims to improve what we can here with minimal changes. A
few things we took advantage of to accomplish this:

- Turning the extracted stack into a tuple so it is hashable so it can be used
  as a dictionary key. This let's us check if the stack is indexed first, and
  skip indexing the frames again. This is especially effective in profiles where
  it's blocking on a network request for example, since there will be many
  identical stacks.
- Using the hash of the stack as the dictionary key. Hashing the entire stack
  can be an expensive operation since a stack can have up to 128 frames. Using
  it as a dictionary key means it needs to be rehashed each time. To avoid this,
  we pre-hash the stack and use the hash as a dictionary key which is more
  efficient.
- Convert numbers to strings ahead of time if we know have to. Values like the
  tid and elapsed since start ns needs to be sent as a string. However, many
  samples share the same value for it, and we're doing the conversion each time.
  Instead, we convert them to a string upfront and reuse it as needed in order
  to minimize unnecessary calculations.

…ion-code-for-performance