bfd/riscv: tighten matching rules in riscv_scan

The following GDB behaviour was observed:

  (gdb) x/1i 0x0001014a
     0x1014a <main+8>:	jal	0x10132 <foo>
  (gdb) show architecture
  The target architecture is set automatically (currently riscv:rv32)
  (gdb) set architecture riscv:rv32
  The target architecture is assumed to be riscv:rv32
  (gdb) x/1i 0x0001014a
     0x1014a <main+8>:	0x37e5
  (gdb)

Notice that initially we can disassemble the instruction (it's a
compressed jal instruction), but after setting the architecture we can
no longer disassemble the instruction.

This is particularly puzzling as GDB initially thought the
architecture was 'riscv:rv32', but when we force the architecture to
be that, the disassembly stops working.

This issue was introduced with this commit:

  commit c35d018b
  Date:   Mon Jan 27 15:19:30 2020 -0800

      RISC-V: Fix gdbserver problem with handling arch strings.

In this commit we try to make riscv_scan handle cases where we see
architecture strings like 'riscv:rv32imc' (for example).  Normally
this wouldn't match as bfd_default_scan requires an exact match, so we
extended riscv_scan to ignore trailing characters.

Unfortunately the default riscv arch is called 'riscv', is 64-bit,
and has its mach type set to 0, which I think is intended to pair with
code is riscv-dis.c:riscv_disassemble_insn that tries to guess if we
are 32 or 64 bit.

What happens then is that 'riscv:rv32' is first tested against 'riscv'
using bfd_default_scan, this doesn't match, we then compare this to
'riscv', but allowing trailing characters to be ignored, this matches,
and our 'riscv:rv32' matches against the default (64-bit)
architecture.

The solution I propose is to prevent the default architecture from
taking part in this "ignore trailing characters" extra match case,
only the more specific 'riscv:rv32' and 'riscv:rv64' get this extra
matching.

bfd/ChangeLog:

	* cpu-riscv.c (riscv_scan): Don't allow shorter matches using the
	default architecture.