With this change, we can now select implementations of most
transcendentals from either the "emulated" or "optimal"
namespaces. The emulated versions generally call the math.h
standard function on each component for vector types, while
the optimal versions typically implement some approximation
in Reactor to produce vectorized code.

Most of the optimal versions were taken directly from ShaderCore.cpp,
except for ASin, for which I implemented an 8-term approximation. The
new versions are faster, and pass all deqp precision tests.

Here's a table of benchmarks that show the performance improvements that
were made. Note that Asin and Acos now take a Precision parameter for
Full and Relaxed precision:

                     Before     After
rr_Sin               48.6 ns    10.6 ns
rr_Cos               67.1 ns    9.62 ns
rr_Tan               75.5 ns    19.4 ns
rr_Asin_fullp        24.2 ns    23.0 ns
rr_Asin_relaxedp     N/A        9.31 ns
rr_Acos_fullp        14.3 ns    6.35 ns
rr_Acos_relaxedp     N/A        4.56 ns
rr_Atan              66.8 ns    12.9 ns
rr_Sinh              79.7 ns    11.5 ns
rr_Cosh              80.1 ns    11.5 ns
rr_Tanh              62.9 ns    12.1 ns
rr_Asinh              104 ns    9.44 ns
rr_Acosh             14.4 ns    10.2 ns
rr_Atanh              170 ns    9.81 ns
rr_Atan2             73.5 ns    22.8 ns
rr_Pow               87.9 ns    16.3 ns
rr_Exp               40.2 ns    5.72 ns
rr_Log               44.0 ns    7.35 ns
rr_Exp2               101 ns    5.38 ns
rr_Log2               106 ns    9.24 ns

Bug: b/147818976
Change-Id: I791893bd9f005dbbae4770fb474de338a04845be
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/48588
Kokoro-Result: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Antonio Maiorano <amaiorano@google.com>