call mrb_close() in tools and test programs

Avoid warning: redundant redeclaration of ‘OnigEncDefaultCharEncoding’

Add CMake build infrastructure

I was hoping this would cause valgrind to complain less, but there is still
a lot of memory leaked (does mrb_close() actually free all of the managed
blocks?)  Anyway this helps somewhat and is good practice

Conflicts:
	src/time.c

As per monaka's comment here...
  https://github.com/mruby/mruby/pull/175
...changing these initializations of mrb_int values to use integer
constants

* mrb_time_gm() uses mrb_int
* Same for mrb_time_initialize()
* The parameters to time_mktime() are mrb_int, and that's what these values are passed to

So I don't see why mrb_time_local() works on mrb_float instead.  Lets make
it consistent with the other methods defined in time.c

More C++ compilability work: mrb_obj_alloc void* conversions

One of the biggest set of changes needed to make C++ compile, is that you
can't autoconvert "void*" to a different pointer type without a cast (you
can of course, convert pointers *to* "void*"!)

For the first part, convert the users of "mrb_obj_alloc".  Since it has
to return something, make it RBasic* (that's what mrb_obj_alloc() is
operating on anyway).  This way, even in C you'll get a warning if you
don't cast it.

For places where there are a lot of similar calls to mrb_obj_alloc(),
this can be easily hidden through a macro.  I did this in string.c:
    #define mrb_obj_alloc_string(mrb) ((struct RString *) mrb_obj_alloc((mrb), MRB_TT_STRING, (mrb)->string_class))

I also updated the mrb_object() macro to also return a RBasic* -- my
previous commit changed that from "void*" -> "RObject*", but I figure
it should be consistent with mrb_obj_alloc()

Part 1 of C++ compilability

Add cast to remove "cast from pointer to integer of different size" warning

Fix Hash#shift return value from Hash to Array

C++ is pickier about when a 'goto' can cross a variable being delcared.
The fix is to just add a set of braces to restrict the variable's scope.
Without this, g++ will fail with:

  regcomp.c:3057: error: jump to label 'set_call_attr'
  regcomp.c:3087: error:   from here
  regcomp.c:3041: error:   skips initialization of 'int gnum'

I originally solved this by renaming the "node" type to "node_t",
but Matz didn't like that.  He suggested renaming the member
variable "nd" instead:
  https://github.com/mruby/mruby/pull/144#issuecomment-5743153

C++ is picker than C about when you can "goto" across a variable being
defined.  The fix is to just minimize the variable's scope inside an
extra set of brackets.

Without this change, g++ has the following errors:

  transcode.c:590: error: jump to label 'resume_label3'
  transcode.c:514: error:   from here
  transcode.c:582: error:   crosses initialization of 'const unsigned char* p'
  transcode.c:2124: error: jump to label 'set_encs'
  transcode.c:2184: error:   from here
  transcode.c:2088: error:   skips initialization of 'const char* err'
  transcode.c:2089: error:   skips initialization of 'size_t error_len'
  transcode.c:2090: error:   skips initialization of 'mrb_value bytes'
  transcode.c:2091: error:   skips initialization of 'mrb_value dumped'
  transcode.c:2092: error:   skips initialization of 'size_t readagain_len'
  transcode.c:2093: error:   skips initialization of 'mrb_value bytes2'

See Matz's comment here:
  https://github.com/mruby/mruby/pull/144#issuecomment-5754054

mrb_gc_mark_ht_size() and mrb_gc_free_ht() were declared in gc.h as
taking a "RHash *" argument, but then they were defined in hash.c
as taking a "RClass *"  Get these in sync.

see http://en.cppreference.com/w/cpp/language/operator_alternative