Smarter `number(float)` buffer

The code currently handles floats by using a 4096 buffer space. We can do better:

- There are 2 main pieces to floats: mantissa + exponent. Both are known to be limited.
- Max length should be max digits of mantissa + max digits of exponent + formatting flags

Thoughts on a limiting parser:
1. Fill up the buffer until mantissa is full
2. If mantissa is full, don't append into buffer. If we didn't see a decimal, keep accruing the digit overflow; this is going to append to the actual exponents.
3. If we find `e`, treat the next set of digits as exponents.
4. Toss this result weird monstrosity result into `parseFloat`

I believe `parseFloat` should handle all of these edge cases already, in which case it could possibly be adapted to understand a `reader` instead of just strings. So maybe we could push an update to stdlib.