Optimize Base.decode*

[sabiwara]

Since #14429 seemed to indicate that using rescue to check for tuple membership was not optimal, I tried to replace it by defensively checking the index (not beautiful I know 😅).
The result looks promising, although since decode_ function is overall doing more work than valid_? the results are less impressive in comparison.

The result is more flaky here - sometimes getting 14% speed increase, sometimes 55% - but it's consistently quite faster for a big string.

##### With input small string #####
Name               ips        average  deviation         median         99th %
no_raise        3.50 M      285.86 ns ±10510.17%         208 ns         375 ns
original        3.32 M      301.58 ns ±10681.55%         208 ns         375 ns

Comparison: 
no_raise        3.50 M
original        3.32 M - 1.05x slower +15.72 ns

Memory usage statistics:

Name        Memory usage
no_raise           168 B
original           168 B - 1.00x memory usage +0 B

**All measurements for memory usage were the same**

##### With input big string #####
Name               ips        average  deviation         median         99th %
original        254.71        3.93 ms     ±6.28%        3.90 ms        4.19 ms
no_raise        205.72        4.86 ms    ±42.61%        3.44 ms        8.25 ms

Comparison: 
original        254.71
no_raise        205.72 - 1.24x slower +0.93 ms

Memory usage statistics:

Name        Memory usage
original           184 B
no_raise           184 B - 1.00x memory usage +0 B

**All measurements for memory usage were the same**

Code

Benchee.run(
  [
    original: &Base.decode64!/1,
    no_raise: &MyBase.decode64!/1
  ],
  memory_time: 2,
  # reduction_time: 1,
  inputs: [
    "small string": Base.encode64("hello world"),
    "big string": Base.encode64(:crypto.strong_rand_bytes(1_000_000))
  ]
)

Also, perhaps we could reduce the duplication here, WDYT?