Improve the performance of layout mapping by explicit loop

Summary

The performance of the tensor access depends on the implementation of the corresponding mapping. In my tests it was better to write a loop to compute the mapping that an innerproduct of slices and extents. This needs to be investigated further. In the libc++ implementation of clang-18 they use a fold-expression of the comma-operator, essentiall beeing equivalent to the loop approach but completely unrolled. This was slightly slower in my test. But I only tested with gcc12. In clang the FieldTensor was at the end slower than a DynamicTensor if the rank is large. This is different in gcc.

Timings

Example: bench_ten_ten with N=4

Clang 18 (libc++)

Benchmark	Time	CPU	Iterations
BM_FieldTensor_FieldTensor	4577 ns	4576 ns	134356
BM_DynamicTensor_DynamicTensor	2634 ns	2634 ns	265871

Clang 18 (index(i)*stride(i) +...+ 0)

Benchmark	Time	CPU	Iterations
BM_FieldTensor_FieldTensor	4514 ns	4514 ns	135101
BM_DynamicTensor_DynamicTensor	3159 ns	3159 ns	222465

Clang 18 (for-loop - this MR)

Benchmark	Time	CPU	Iterations
BM_FieldTensor_FieldTensor	4524 ns	4523 ns	135125
BM_DynamicTensor_DynamicTensor	2608 ns	2608 ns	267791

GCC-12 (index(i)*stride(i) +...+ 0)

Benchmark	Time	CPU	Iterations
BM_FieldTensor_FieldTensor	1172 ns	1172 ns	588238
BM_DynamicTensor_DynamicTensor	5996 ns	5995 ns	115243

GCC-12 (for-loop - this MR)

Benchmark	Time	CPU	Iterations
BM_FieldTensor_FieldTensor	1140 ns	1140 ns	603939
BM_DynamicTensor_DynamicTensor	2931 ns	2931 ns	239035