Improvements to forward triplet search

This MR adds a linear search and optimizes the forward triplet search algorithm a bit.

• linear_search is a new algorithm that takes a similar input to binary_search_leftmost. It accepts an optional starting index, from which the array is sought in either increasing or decreasing order as required. It outperforms binary search in specific circumstances.

lf_triplet_seeding got the following improvements:

• It now uses a shared memory array shared_xs, which holds all x coordinate values of the three modules in consideration.
• It additionally uses shared_indices and shared_number_of_elements to store x0 and x2 indices that passed the cut.
• Adding the hit in the middle module x1 is done afterwards by iterating shared_indices, improving thread usage.
• syncwarp instead of syncthread are used. It is more technically correct, although it doesn't seem to have an impact on performance.

Physics efficiency is expected to improve ever-so-slightly due to the better load balancing of candidate analysis between threads.

• Throughput change: 162.35 kHz -> 165.71 kHz (1.02x)
• Physics change: https://gitlab.cern.ch/lhcb/Allen/-/jobs/9592025

Built on top of !414 (merged)

backend/include/CPUBackend.h

@@ -21,6 +21,7 @@ using std::signbit;
 #define __global__
 #define __constant__
 #define __syncthreads()
+#define __syncwarp()
 #define __launch_bounds__(_i)
 #define cudaError_t int
 #define cudaEvent_t int