A smaller, yet key, step to move

From: a situation where mainly executables (but not really) depend on
allocator, and base needs dependencies (to tcmalloc) to be injected
from content (which violates the ODR in component buids).

To: a situation where only base depends on allocator and the other
targets get recursively the required linked flags.

In essence this CL is a more gradual approach to the bigger
unreviewable crrev.com/1528013002.

How is the transition handled?
------------------------------
After this CL, the situation will be as follows:
From a build time perspective base will also depend on allocator.
This will not change anything substantial in static builds and introduce
yet another (temporary) ODR violation in Linux component builds.
The big change introduced by this CL is the fact that all the executable
targets that depend on base (virtualy all) will also get another
indirect dependency to allocator.

In other words, after this CL executable targets will depend on
allocator for two reasons:
 - Because they have an explicit dependency to it (the one I am going to
   get rid of in the immediate future).
 - Because this new transitive path I am introducing in base.

Rationale of this approach
--------------------------
This allows to restrict the critical changes in a smaller CL easier to
review, at the cost of the temporary double dependency on base.
The good things are:
 - If something will break, this CL will be very easy to revert.
 - The next cleanups will be straightforward.
 - We have now smoke tests (crrev.com/1577883002) that will help us
   realize if something goes wrong.

Next steps
----------
In the next CLs I will:
 - Remove the content -> base injection layer, and let base directly use
   the tcmalloc functions it needs.
 - Remove all the traces of USE_TCMALLOC outside of base.
 - Start cleaning up the hundreds use_allocator conditionals in the gyp
   files in a way which is easier to review and produce zero ninja diffs
   (see crrev.com/1583973002 as an example)

Ninja diffs caused by this change
---------------------------------
 ### Win, static build, GN:   https://paste.ee/p/hvcRp
 The missing targets (mostly tests) that previously were
 not depending on allocator, now get that by virtue of the transitive
 dependency.

 ### Win, static build, GYP:  https://paste.ee/p/AGuKR
 As above. Just GYP seems to emit the ninja files in a different,
 inlined, format.

 ### linux static build, GYP:  https://paste.ee/p/kmD7U
 As above. Plus the new targets also get the -Wl,-u (keep symbol)
 args as expected by allocator.gyp for the tcmalloc heap profiler.

 ### linux shared build, GYP: https://paste.ee/p/FHHNR
 Nothing relevant. Just I moved the dependency to allocator from
 base_unittests to base, and that is the only thing that reflects in the
 ninja files.

BUG=564618

Review URL: https://codereview.chromium.org/1584893002

Cr-Commit-Position: refs/heads/master@{#370405}