fix: distinguish a stored nil from a missing key in fetch/2

[nmbrone]

Because fetch/2 was implemented on top of get/3, a key holding nil was reported as :error — indistinguishable from a key that was never set:

Rkv.put(:bucket, :key, nil)
Rkv.fetch(:bucket, :key)
#=> :error   # expected {:ok, nil}

That defeats the only reason fetch/2 exists alongside get/3.

fetch/2 now reads from ETS directly and get/3 is defined in terms of it, so there is a single lookup site. Cost is unchanged: one registry read for the table id, one :ets.lookup/2.

Covered by a test asserting a stored nil and a missing key are told apart.