Improve runtime performance - lazy stringification, more inlining, no statics on the hot path, move semantics for classes such as doctest::String which are used by value, etc.

[onqtam]

No description provided.

[onqtam]

The following code:

TEST_CASE("") {
    for(int i = 0; i < 10000000; ++i)
        CHECK(i == i);
}

used to run for 9.0 seconds built with MinGW-w64 g++ 7.1 in Release.

Now it runs for 0.27 seconds.

There were 3 changes that I introduced:

• A new implementation of the String class - now with a small buffer optimization - sizeof(String) is 24 and it can hold strings of length up to 23 characters (24 with the null terminator) in itself!
  The trick used for this maximum length on the stack is explained in this cppcon talk about how fbstring (from facebook/folly) is implemented.
  This got down the runtime to 4.1 seconds for that piece of code
• Made the asserts to stringify the expressions lazily - only if they fail.
  This got down the runtime to 0.275 seconds for that code.
• Reworked the access to the current context state to access a global instead of a local static - tiny improvement - perhaps 2-3%
  ~0.27 after this optimization

So this all amounts to a x33 faster runtime execution of asserts!

There are >> 0 << allocations while looping the assert.

All asserts benefit from these optimizations including the binary ones and the fast ones (they are the fastest not only for compile times but also for runtime) - they might not be 33 times faster but will still be a lot faster than before.

Results are similar with other compilers - will update the benchmark part of the documentation soon.