Relax decode bounds check

[wemeetagain]

In #20, a check was made to prevent maliciously crafted input from resulting in NaN or Infinity decoded values.

The check only allows varints with bitlength <= 48 to be decoded.
This is strict, perhaps overly strict. It disallows this library to decode varint64, an extremely common usecase.

But this leads us to a difficult situation without a deeper and more invasive code change -- a tradeoff between allowing a useful range of input buffer lengths and ensuring correctness of all outputs:

1. If we disallow varints with bitlength > 49, we cannot support uvarint64, but we can guarantee the correctness of all outputs (as we simply disallow decoding values > Number.MAX_SAFE_INTEGER).
2. If we allow varints with bitlength > 49, we can support uvarint64, but we cannot guarantee the correctness of the output if the decoded value is > Number.MAX_SAFE_INTEGER.

#20, was written with the intent of 1.

Talking with other teams who use this library, it seems that 2. is the more practical, useful.
cc @vasco-santos

This PR keeps the core check of #20, but relaxes it to allow for varint64.

[vasco-santos]

LGTM!
Thanks for having a look @wemeetagain