inflate: Keep dict on stack (#581)

```
λ benchcmp before.txt  after.txt
benchmark                               old ns/op     new ns/op     delta
BenchmarkDecodeDigitsSpeed1e4-32        45478         44642         -1.84%
BenchmarkDecodeDigitsSpeed1e5-32        520045        506007        -2.70%
BenchmarkDecodeDigitsSpeed1e6-32        5213200       5067185       -2.80%
BenchmarkDecodeDigitsDefault1e4-32      49476         48936         -1.09%
BenchmarkDecodeDigitsDefault1e5-32      496733        490872        -1.18%
BenchmarkDecodeDigitsDefault1e6-32      4869294       4851319       -0.37%
BenchmarkDecodeDigitsCompress1e4-32     44262         44419         +0.35%
BenchmarkDecodeDigitsCompress1e5-32     470795        473693        +0.62%
BenchmarkDecodeDigitsCompress1e6-32     4690131       4689175       -0.02%
BenchmarkDecodeTwainSpeed1e4-32         49316         49218         -0.20%
BenchmarkDecodeTwainSpeed1e5-32         531644        527865        -0.71%
BenchmarkDecodeTwainSpeed1e6-32         5316889       5281027       -0.67%
BenchmarkDecodeTwainDefault1e4-32       49570         49543         -0.05%
BenchmarkDecodeTwainDefault1e5-32       492474        488411        -0.83%
BenchmarkDecodeTwainDefault1e6-32       4902696       4844194       -1.19%
BenchmarkDecodeTwainCompress1e4-32      47788         47235         -1.16%
BenchmarkDecodeTwainCompress1e5-32      465616        454616        -2.36%
BenchmarkDecodeTwainCompress1e6-32      4606437       4513280       -2.02%
BenchmarkDecodeRandomSpeed1e4-32        299           296           -0.94%
BenchmarkDecodeRandomSpeed1e5-32        1930          1942          +0.62%
BenchmarkDecodeRandomSpeed1e6-32        19980         19990         +0.05%
```

flate/_gen/gen_inflate.go

@@ -48,7 +48,7 @@ func (f *decompressor) $FUNCNAME$() {
 	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
 	// but is smart enough to keep local variables in registers, so use nb and b,
 	// inline call to moreBits and reassign b,nb back to f on return.
-	fnb, fb := f.nb, f.b
+	fnb, fb, dict := f.nb, f.b, &f.dict
 
 	switch f.stepState {
 	case stateInit:
@@ -106,9 +106,9 @@ readLiteral:
 		var length int
 		switch {
 		case v < 256:
-			f.dict.writeByte(byte(v))
-			if f.dict.availWrite() == 0 {
-				f.toRead = f.dict.readFlush()
+			dict.writeByte(byte(v))
+			if dict.availWrite() == 0 {
+				f.toRead = dict.readFlush()
 				f.step = (*decompressor).$FUNCNAME$
 				f.stepState = stateInit
 				f.b, f.nb = fb, fnb
@@ -251,10 +251,10 @@ readLiteral:
 		}
 
 		// No check on length; encoding can be prescient.
-		if dist > uint32(f.dict.histSize()) {
+		if dist > uint32(dict.histSize()) {
 			f.b, f.nb = fb, fnb
 			if debugDecode {
-				fmt.Println("dist > f.dict.histSize():", dist, f.dict.histSize())
+				fmt.Println("dist > dict.histSize():", dist, dict.histSize())
 			}
 			f.err = CorruptInputError(f.roffset)
 			return
@@ -267,14 +267,14 @@ readLiteral:
 copyHistory:
 	// Perform a backwards copy according to RFC section 3.2.3.
 	{
-		cnt := f.dict.tryWriteCopy(f.copyDist, f.copyLen)
+		cnt := dict.tryWriteCopy(f.copyDist, f.copyLen)
 		if cnt == 0 {
-			cnt = f.dict.writeCopy(f.copyDist, f.copyLen)
+			cnt = dict.writeCopy(f.copyDist, f.copyLen)
 		}
 		f.copyLen -= cnt
 
-		if f.dict.availWrite() == 0 || f.copyLen > 0 {
-			f.toRead = f.dict.readFlush()
+		if dict.availWrite() == 0 || f.copyLen > 0 {
+			f.toRead = dict.readFlush()
 			f.step = (*decompressor).$FUNCNAME$ // We need to continue this work
 			f.stepState = stateDict
 			f.b, f.nb = fb, fnb