multi optimization make torch_ccl more friendly

[KimBioInfoStudio]

• replace multiprocessing.cpu_count with os.cpu_count()
• replace self defined which func with pyhton builtin func shutil.which()
• check cxx compiler flag before use -fsycl when xpu enabled
• raise friendly error when import torh_ccl from it's source root path

[chengjunlu]

@KimBioInfoStudio Thank you for your effort to make torch-ccl more user friendly. All the changes are reasonable and clean.