Renderer#escape was each time passing a new hash instance to gsub,
making it allocate more memory than neccessary. Moving the hash to a
constant improves performance a bit.

For `Markd.to_html("hello world")`

before 350.46k (  2.85µs) (± 6.31%)  4.09kB/op
after  431.05k (  2.32µs) (± 1.43%)  3.07kB/op

For `Markd.to_html("hello world")`

before  431.05k (  2.32µs) (± 1.43%)  3.07kB/op
after   506.43k (  1.97µs) (± 2.50%)  2.72kB/op

For `Markd.to_html("hello world")`

before 506.43k (  1.97µs) (± 2.50%)  2.72kB/op
after  554.18k (  1.80µs) (± 1.10%)  2.41kB/op

For `Markd.to_html("hello world")`:

before 554.18k (  1.80µs) (± 1.10%)  2.41kB/op
after  640.45k (  1.56µs) (± 6.70%)  2.06kB/op

This avoids initializing a Hash when not needed.

For `Markd.to_html("hello world")`:

before 640.45k (  1.56µs) (± 6.70%)  2.06kB/op
after  727.39k (  1.37µs) (± 2.28%)  1.59kB/op

Array literals in tight loops allocate a lot of memory!

Parsing markd's README.md:

before 3.52k (283.87µs) (± 0.59%)  259kB/op
after  3.66k (273.38µs) (± 0.54%)  244kB/op

For `Markd.to_html("hello world")`:

before 726.36k (  1.38µs) (± 1.86%)  1.59kB/op
after  800.81k (  1.25µs) (± 0.58%)  1.58kB/op

Markdown is an ASCII format: all control characters are ASCII so we
can avoid decoding unicode characters when not needed. The old code was
indexing bytes by unicode index position which is very slow because the
string must be decoded each time. Here we change it to index by byte,
and change all operations to do byte slices instead of unicode slices.

For markd's README.md:

before 3.86k (258.86µs) (± 0.52%)  242kB/op
after  4.11k (243.25µs) (± 1.37%)  242kB/op

For `Markd.to_html("hello world")`:

before 800.81k (  1.25µs) (± 0.58%)  1.58kB/op
after  829.33k (  1.21µs) (±25.35%)  1.56kB/op

We can use an inline comparison or a tuple.

For markd's README.md:

before 4.57k (218.68µs) (± 0.64%)  198kB/op
after  4.63k (215.84µs) (± 2.20%)  196kB/op

`values.each` will create an intermediary array

We can use `each_line` instead of `lines` and then iterating.

For `Markd.to_html("hello world")`:

before 882.84k (  1.13µs) (± 1.46%)  1.56kB/op
after  929.11k (  1.08µs) (± 2.75%)  1.46kB/op

For markd's README.md:

before 4.85k (206.37µs) (± 2.24%)  187kB/op
after  4.98k (200.82µs) (± 0.47%)  183kB/op

`String::Builder` is slightly faster when used for a one-time shot.

Because all escape chars are ASCII it's fine if we go byte per byte.

For `Markd.to_html("hello world")`:

before 942.28k (  1.06µs) (± 5.45%)  1.46kB/op
after  959.11k (  1.04µs) (± 0.72%)  1.46kB/op

For markd's README.md:

before 4.97k (201.41µs) (± 0.74%)  176kB/op
after  5.05k (197.95µs) (± 1.51%)  176kB/op

For markd's README.md:

before 5.14k (194.65µs) (± 1.25%)  174kB/op
after  5.23k (191.14µs) (± 0.55%)  173kB/op