ecpg: fix ECPGstore_input()

This patch fixes the following issues in ECPGstore_input():

- strlen() was invoked on the NULL pointer for the first iteration of
the loop (line 875, 923, 966, 1009)

- `nval' is freed for every iteration of the loop at 864, but only
initialized once outside the loop, resulting in potential multiple
free()'s, as well as the use of a freed variable in subsequent iterations

- `str' was leaked for every subsequent iteration of the loop (line 871,
920, 963, 1006)

- the return value of PGTYPESinterval_to_asc() is leaked at line 920 and
937; the return value of PGTYPESdate_to_asc() is leaked at line 963 and
980; the return value of PGTYPEStimestamp_to_asc() is leaked at line
1006 and 1023.

- malloc failure is in general not handled well; the function simply
returns without bothering to clean up allocated resources, and many
return values are not checked for errors.

Also, in create_statement(), `*stmt' was dereferenced before being
initialized.

Per the Coverity report run by EnterpriseDB. Thanks to Eric Astor at EDB
for an initial version of this patch -- the attached version has been
improved by myself.

Barring any objections, I'd like to apply this to CVS in a day or two (I
want to test it first, which I haven't yet done).

-Neil