WL#6325: Refactoring: no more half-initialized SELECT_LEX

- a LEX contains a default SELECT_LEX (lex->select_lex)
- a LEX has a pointer current_select to the current SELECT_LEX
- some statements (LOAD DATA, ALTER TABLE, ...) have only the default
SELECT_LEX, whereas others (SELECT, ...) create a new one in mem_root.
- lex_start() (called for every statement) initializes a bit of thd->lex->select_lex
- lex_start() calls st_select_lex::init_query() which also initializes a bit of
a SELECT_LEX
- both inits above thus look like "what's needed for all queries"
- st_select_lex::init_select() initializes the rest of the SELECT_LEX but not
all statements call it (looks like "what's additionally needed for SELECT-like
queries")
- Some statements (like DELETE) do access a SELECT_LEX which has only gone
through st_select_lex::init_query() (DELETE stores its table in
current_select->tables_list).

Thus, inside other code it is hard to know if certain members of
thd->lex->current_select are initialized or not. Example:
see guilhem.bichot@oracle.com-20120529155612-l84yb1x7oi0v1q7w in trunk.

It is bad design for a class (LEX) to let other pieces of code
access a half-initialized object. If *(lex->current_select) is random
then lex->current_select should not point there. It should rather
point to NULL, and be set to non-NULL only when the SELECT_LEX is
fully ready.

Proposal:
- decide if we really have to skip st_select_lex::init_select() for
some statements (i.e. produce a half-initialized object) (the reason
might be performance)
- if no, move st_select_lex::init_select() into st_select_lex::init_query()
- if yes, split the object in two sub-objects (what's for all queries, what's
for SELECT-like)
- if something is not reliable, don't point to it; rather use a NULL
pointer, so the rest of code knows that there is nothing to be found
there.

See also WL#4772.