Suppress unrolling by clang

[nomaddo]

Currently the front-end clang unroll loops because -O3 is specified as compilation option.
This makes harder to do

• fusion of write back to memory
• vectorization in VC4C unrolling
• Suppress explosion in use of registers

It requires careful optimization design, but tt's better to do in our optimization pass.

[doe300]

I'm not sure about this one. Sure, if we have full control over loop unrolling, we can adapt it more to our platform. But I would not trust the efficiency and correctness of out optimizations over the clang compiler;)

[nomaddo]

Unrolling inner-loops is easy compared with other loop optimizations.
Vectorization require some analysis no to make wrong program, but unrolling inner-loops are always correct.