kernel/fcall.c optimizations

[ghost]

• renamed all macros to follow the same style: phalcon_call_WHAT_pX(_noret)
• fixed inconsistencies in restoring memory frame on failure in phalcon_create_instance(_params)
• use variadic arguments instead of arrays in phalcon_call_WHAT_params()
• added phalcon_has_constructor_ce() which reuses the existing class entry (calls to Z_OBJCE_P() are expensive)
• removed a lot of functions from kernel/fcall.h — now that we don't need array to pass the parameters, we can use variadic macros instead
• use do_alloca()/free_alloca() macros from Zend instead of emalloc()/efree(): if the block to be allocated is not large, alloca() will be used (much faster than emalloc()); otherwise, emalloc() will be used as a fallback
• introduced optimized macros PHALCON_CALL_FUNCTION, PHALCON_CALL_METHOD, PHALCON_CALL_PARENT, PHALCON_CALL_SELF, PHALCON_CALL_STATIC: they wrap phalcon_call_XXX_params() functions and restore memory frame on failures. If the compiler is GCC, these macros expand to pretty optimized calls: when the method or function name is a known compile time constant, its hash can also be computed at compile time. If the name is not a literal (say, a variable), then instead of sizeof() we use strlen() to make sure that the length will be correct.

PS: I will address failed Mongo tests in my next PR as they are not related to fcall.

[ghost]

Just added:

• when method hash is not supplied, check if the method name is an interned string and if so, use the precomputed hash