Pool allocate ImplicitConversionSequences.

To avoid malloc thrashing give OverloadCandidateSet an inline capacity for conversion sequences.
We use the fact that OverloadCandidates never outlive the OverloadCandidateSet and have a fixed
amount of conversion sequences.

This eliminates the oversized SmallVector from OverloadCandidate shrinking it from 752 to 208 bytes.

On the test case from the "Why is CLANG++ so freaking slow" thread on llvmdev this avoids one gig
of vector reallocation (including memcpy) which translates into 5-10% speedup on Lion/x86_64.

Overload candidate computation is still the biggest malloc contributor when compiling templated
c++ code.

llvm-svn: 148186