Access @rs only through attr_accessor

Ruby does not pass the nesting to const_missing (unfortunately).
That second argument was there in case that changed, Yehuda
sent a patch to MRI

    http://bugs.ruby-lang.org/issues/2740

but there is not much movement there and Matz told me in
Amsterdam there was no immediate plan to pass the nesting.
So let's go back to implement what happens now, and if
in the future we get the nesting then we will adapt this.

Double-checked this with Mr Katz.

This method was added to be shared between update_attribute and
update_column in 50725cec, but since
update_attribute was removed, and update_column has changed to delegate
to update_columns, the method is not used anywhere anymore.

Also remove "key.to_s" conversion when raising readonly error, since
the key is being interpolated.

Fix "last equality wins" logic in relation merge

This is a real fix (as compared to the band-aid in b127d86c), which uses
the recently-added equality methods for ARel nodes. It has the side
benefit of simplifying the merge code a bit.

Use inversed parent for first and last child of has_many association

Added X-Content-Type-Options to the header defaults.

With a value of "nosniff", this prevents Internet Explorer from MIME-sniffing a response away from the declared content-type.

Refactor multiparameter attributes assignment implementation.

Cleans up a lot of noise from arguments being passed from one method to
another.

This should make it easier to refactor and improve this code, and remove
complexity with params going around here and there.

Avoid doing a new column lookup for the attribute, since we already have
the column to check for the klass.

Move some methods to the top to better organize them, since they're used
right at the beginning of the multiparameter assignment method chain.

Replace deprecated `memcache-client` gem with `dalli` in ActiveSupport::Cache::MemCacheStore

memcache-client was deprecated in favour of dalli in 2010.

Fix merge error when Equality LHS is non-attribute

Update to activerecord-deprecated_finders on generator

Checking respond_to? incurs overhead, and most of the time when
assigning attributes it will return true. So just handle the
NoMethodError instead.

any? will check that each item in the array is truthy, as opposed to
!empty? which will simply check that the array has length. For an empty
array, !empty? still seems to be faster than any?

Rather than doing it every time an instance is instantiated.

Those z's were hard to type.

deep_dup is slow. we only need to dup the values, so just do that
directly.

Fixes #7374

Fix occasional microsecond conversion inaccuracy

This is at best a band-aid for a more proper fix, since it won't truly
handle the removal of the previous equality condition of these other
nodes. I'm planning to put in some work on ARel toward supporting that
goal.

Related: rails/arel#130, ernie/squeel#153, ernie/squeel#156